HR001119S0085.pdf

Broad Agency Announcement Semantic Forensics (SemaFor)

HR001119S0085

August 23, 2019

Defense Advanced Research Projects Agency Information Innovation Office 675 North Randolph Street Arlington, VA 22203-2114

HR001119S0085 SEMAFOR 2

Table of Contents

Part I: Overview Information……………………………………………………………………...………3

Part II: Full Text of Announcement………………………………………………………………………..4

I. Funding Opportunity Description

II. Award Information

A. Awards

B. Fundamental Research

C. Disclosure of Information and Compliance with Safeguarding Covered Defense Information Controls

III. Eligibility Information

A. Eligible Applicants

B. Organizational Conflicts of Interest

C. Cost Sharing/Matching

D. Other Eligibility Requirements

IV. Application and Submission Information

A. Address to Request Application Package

B. Content and Form of Application Submission

C. Submission Dates and Times

D. Funding Restrictions

E. Other Submission Requirements

V. Application Review Information

A. Evaluation Criteria

B. Review and Selection Process

VI. Award Administration Information

A. Selection Notices

B. Administrative and National Policy Requirements

C. Reporting

VII. Agency Contacts

VIII. Other Information

A. Frequently Asked Questions (FAQs)

B. Proposers Day

HR001119S0085 SEMAFOR 3

C. Submission Checklist

D. Associate Contractor Agreement (ACA)

HR001119S0085 SEMAFOR 4

PART I: OVERVIEW INFORMATION

Federal Agency Name: Defense Advanced Research Projects Agency (DARPA), Information Innovation Office (I2O)

Funding Opportunity Title: Semantic Forensics (SemaFor)

Announcement Type: Initial Announcement

Funding Opportunity Number: HR001119S0085

Catalog of Federal Domestic Assistance Numbers (CFDA): 12.910 Research and Technology Development

Dates o Posting Date: August 23, 2019 o Abstract Due Date: September 11, 2019, 12:00 noon (ET) o Proposal Due Date: November 21, 2019, 12:00 noon (ET) o Proposers Day: August 28, 2019

Anticipated Individual Awards: DARPA anticipates multiple awards for Technical Areas 1 and 4, and single awards for Technical Area 2 and 3.

Types of Instruments that May be Awarded: Procurement contracts, grants, cooperative agreements, or Other Transactions

Agency Contacts o Technical POC: Matt Turek, Program Manager, DARPA/I2O o BAA Email: SemaFor@darpa.mil o BAA Mailing Address:

DARPA/I2O

ATTN: HR001119S0085

675 North Randolph Street Arlington, VA 22203-2114 o I2O Solicitation Website: http://www.darpa.mil/work-with-us/opportunities http://www.darpa.mil/work-with-us/opportunities

HR001119S0085 SEMAFOR 5

PART II: FULL TEXT OF ANNOUNCEMENT

I. Funding Opportunity Description

DARPA is soliciting innovative research proposals in the area of semantic technologies to automatically assess falsified media. Proposed research should investigate innovative approaches that enable revolutionary advances in science, devices, or systems. Specifically excluded is research that primarily results in evolutionary improvements to the existing state of practice.

This Broad Agency Announcement (BAA) is being issued, and any resultant selection will be made, using procedures under Federal Acquisition Regulation (FAR) 6.102(d)(2) and 35.016.

Any negotiations and/or awards will use procedures under FAR 15.4 (or 32 CFR § 200.203 for grants and cooperative agreements). Proposals received as a result of this BAA shall be evaluated in accordance with evaluation criteria specified herein through a scientific review process.

DARPA BAAs are posted on the Federal Business Opportunities (FBO) website (https://www.fbo.gov/) and the Grants.gov website (https://www.grants.gov/).

The following information is for those wishing to respond to this BAA.

Introduction/Background

The Semantic Forensics (SemaFor) program will develop technologies to automatically detect, attribute, and characterize falsified, multi-modal media assets (e.g., text, audio, image, video) to defend against large-scale, automated disinformation attacks.

Statistical detection techniques have been successful, but media generation and manipulation technology are advancing rapidly. Purely statistical detection methods are quickly becoming insufficient for detecting falsified media assets. Detection techniques that rely on statistical fingerprints can often be fooled with limited additional resources (algorithm development, data, or compute). However, existing automated media generation and manipulation algorithms are heavily reliant on purely data driven approaches and are prone to making semantic errors. For example, GAN-generated faces may have semantic inconsistencies such as mismatched earrings. These semantic failures provide an opportunity for defenders to gain an asymmetric advantage. A comprehensive suite of semantic inconsistency detectors would dramatically increase the burden on media falsifiers, requiring the creators of falsified media to get every semantic detail correct, while defenders only need to find one, or a very few, inconsistencies.

SemaFor seeks to develop innovative semantic technologies for analyzing media. Semantic detection algorithms will determine if multi-modal media assets have been generated or manipulated. Attribution algorithms will infer if multi-modal media originates from a particular organization or individual. Characterization algorithms will reason about whether multi-modal https://www.fbo.gov/

HR001119S0085 SEMAFOR 6

media was generated or manipulated for malicious purposes. These SemaFor technologies will help identify, deter, and understand adversary disinformation campaigns.

Program Description

DARPA is seeking revolutionary ideas that lead to rigorous, practical demonstrations of the ability to detect, attribute, and characterize falsified multi-modal media. SemaFor will develop methods that exploit semantic inconsistencies in falsified media to perform these tasks across media modalities and at scale. Regular evaluations will target 1000s of multi-modal media assets (e.g., images, text, audio clips, videos), but performers should develop techniques that could be scaled to address internet volumes of media. Focused challenge problems, based on observed trends in media manipulation, will be used over the course of the program to ensure that program capabilities match potential threats. SemaFor methods and the SemaFor system will be expected to operate over increasingly complex media, including reasoning across multiple media assets, while improving detection, attribution, and characterization performance over the duration of the effort.

SemaFor performance will be evaluated on collections of media assets gathered from open sources, as well as multi-modal media assets specifically designed for SemaFor evaluations.

Experiments will be designed to evaluate how well performer algorithms achieve the three main tasks (detect, attribute, and characterize), and will compare performance to human baselines. The purpose of the evaluations is twofold: first, to establish rigorous scientific protocols for measuring the performance of algorithms that reason about potentially falsified media and second, to assess the performance of a SemaFor system in realistic, operational environments.

Proposals describing approaches that rely exclusively on statistical fingerprints, such as sensor noise patterns, are not of interest and are considered out of scope for the program. Also not of interest for SemaFor are TA1, TA2, or TA3 proposals that focus only on a single media modality.

The following terms will be used throughout the BAA, and these definitions are offered to enable clearer communications:

Media modalities: Media forms, examples including text, image, video, and audio.

Media asset: A media instance, such as a single media item and modality: an image, a video, an audio, or text document.

Multi-modal asset: A media collection that may be treated as a single event or instance, such as a news story. May contain some combination of multiple modalities such as image, video, audio, and text.

News articles: A journalist-written story describing an event of interest using multiple modalities. For example, a web page with text and images or video describing an event of interest. News articles are expected to include source organization, an author, and date/time. Some stories may include a location.

Social media post: A short, multi-modal media asset, such as Twitter. Social media posts are expected to be shorter and more colloquial than news articles. Social media posts are expected to include a source platform, an author, and date/time. Depending on

HR001119S0085 SEMAFOR 7

social media type (real or generated) they may provide access to the social network of users.

Technical information: A news story, social media post, or technical article describing a technical capability. For example, a news article describing a new ballistic missile capability.

News collection: Multiple news articles describing a single event. Assets will be from approximately the same time period (e.g., hours to a few days).

Technical information collection: Multiple technical information assets. Assets will be from approximately the same time period (e.g., hours to a few days).

Falsified media: Media that has been manipulated or generated.

Malicious intent: In the context of SemaFor, this relates to media that has been falsified to create a negative real-world reaction. For example, falsifying a story to increase its polarization and likelihood to go viral.

Media source: Purported organization that created a media asset (e.g., a newspaper or news channel).

Media author: Purported individual that created a media asset (e.g., the author, actor, photographer, videographer).

Program Structure

SemaFor has four technical areas (TAs), as shown in Figure 1.

TA1 is Detection, Attribution, and Characterization of multi-modal media assets.

Multiple TA1 awards are expected.

TA2, Explanation and Integration, will combine results across TA1 performers, present explanations for system decisions, and prioritize assets for analyst review. TA2 will also develop the prototype SemaFor system. A single TA2 award is expected.

TA3, Evaluation, will generate and curate multi-modal media for evaluation, design the program evaluations, establish human baseline performance, and develop additional program metrics. A single TA3 award is expected.

TA4, Challenge Curation, will continually provide state-of-the-art (SOTA) challenges to the program to test the techniques developed by TA1 and TA2. TA4 will also develop forward-looking threat models, to anticipate future threats and ensure SemaFor defenses are focused in the right areas. Multiple TA4 awards are expected.

HR001119S0085 SEMAFOR 8

Figure 1: SemaFor Technical Areas

DARPA intends SemaFor to be a collaborative program in which all performers (within and across Technical Areas) constructively interact with one another. To facilitate the open exchange of information, DARPA intends to include associate contractor agreement (ACA) clause in performer awards (see Section VIII.D). This clause is intended to ensure appropriate coordination and potential integration of work done by the SemaFor performers. Once selections have been made, performers should have their ACAs in place prior to the program kick-off meeting.

A key goal of the program is to establish an open, standards-based, multisource, plug-and-play architecture that allows for interoperability and integration. This goal includes the ability to easily add, remove, substitute, and modify software and hardware components in order to facilitate rapid innovation by future developers and users. Therefore, DARPA desires that all software (including source code), software documentation, hardware designs and documentation, and technical data generated by the program be provided as deliverables to the Government as open source software, as lesser rights may adversely affect the lifecycle costs of affected items, components, or processes.

DARPA intends that the development of the SemaFor platform will be driven in part by the creation of a SemaFor data corpus that includes both high integrity (e.g., original) and falsified multi-modal media for development and testing purposes. Due to the highly dynamic/reactive/adversarial nature of media manipulation, DARPA anticipates that the quantity and diversity of data made available may be useful for evaluation and system development, but may not be sufficient to support training, in the sense that has become standard in the computer vision and machine learning communities. Therefore, proposed approaches requiring substantial amounts of training data should describe how sufficient training data of adequate quality will be acquired.

HR001119S0085 SEMAFOR 9

SemaFor will promote broad research community and transition partner involvement by conducting evaluations, periodic challenges, and hackathons as described in this document and shown in the schedule in Figure 3. All performers will be expected to participate in these evaluations as a program requirement. DARPA is not anticipating structured down selects across the phases of the program, although performers may be terminated if the Government deems the research approach is unlikely to yield adequate results.

Technical Areas

Proposers may submit proposals to all TAs. However, each proposal may only address one TA.

DARPA will not make TA1 and TA2 awards to the same performer. The TA3 performer may not perform on TA1 or TA2 due to an inherent conflict of interest with the evaluation process. TA4 performers may be awarded contracts on other TAs of the program, but conflicts of interest plans will be required in the case of TA1 or TA2 due to potential conflicts of interest with the evaluation process. Proposers interested in multiple TAs must submit separate proposals for each TA. In the event that multiple proposals are deemed selectable, the Government reserves the right to choose which to fund, in accordance with the conflict of interest rules described above.

TA1 Detection, Attribution, Characterization

Multiple TA1 teams will focus on the detection, attribution, and characterization (DAC) of falsified multi-modal media. Detection algorithms will examine single and multi-modal media assets, and reason about semantic inconsistencies to determine if the media has been falsified.

Attribution algorithms will analyze the content of media assets with respect to a purported source to determine if the purported source is correct. Attribution algorithms which can also support attributing falsified media to a falsifier organization or individual are also of interest, but not a primary focus. Characterization algorithms will examine the content of a media asset to determine if it was falsified with malicious intent, for example, to significantly alter its tone, polarization, content, or real-world impact.

Methods proposed must go beyond current state of the art techniques while not sacrificing current performance capabilities on single media modalities. Techniques must leverage the most effective statistical detection methods in conjunction with semantic inconsistency detectors created for the SemaFor program. TA1 approaches must handle all media modalities (e.g., text, image, video, audio) for the DAC tasks.

TA1 proposals should describe specifically the types of semantic properties that their algorithms will use as part of the DAC process and the possible limitations of those semantic properties. TA1 efforts likely will need outside semantic knowledge as elements of their DAC process. TA1 proposals should explain the types of outside knowledge that will be incorporated into their algorithms and processing, and how they will scale up their approach to acquiring such knowledge for real-world problems. To help plan program evaluations and to illustrate the utility of a performer’s TA1 approach, TA1 proposals should highlight datasets and scenarios that are aligned with the proposed approach and would provide strong demonstrations of algorithm performance.

HR001119S0085 SEMAFOR 10

TA1 DAC components will provide separate detection, attribution, and characterization scores and evidence (as to why the algorithm produced those scores) to TA2 via SemaFor APIs. DARPA expects TA1 performers will calibrate their scores across performers so that scores from different DAC components are comparable. TA1 and TA2 performers will collaborate to determine a calibration process.

TA1 proposals should describe the types of evidence their DAC algorithms will provide to the TA2 component. Algorithms that perform the DAC tasks but cannot provide an end user with evidence for why a conclusion was reached are of limited value. TA1 components should enable users to understand the basis of system decisions, including what parts of the media asset under analysis may have been falsified.

TA1 performers will collaborate with other TA1 teams and TA2 to design SemaFor system APIs that support the development and evaluation of a scalable SemaFor system. TA1 performers will create implementations of their algorithms that conform to the SemaFor system API, and deliver those implementations and documentation to DARPA and to TA2 for integration into the SemaFor system. DARPA expects that the system design will use Docker containers, message passing frameworks, RESTful design, or similar technologies to manage dependencies and enable rapid system deployment in cloud-based environments. TA1 teams will collaborate with TA2 to ensure software component integration. Evaluations of TA1 components will be done via the integrated SemaFor system.

TA1 performers will also be expected to collaborate closely with the TA2 performers who will be developing score fusion, explanation generation, and media asset prioritization algorithms.

In support of TA2’s responsibilities, TA1 components will need to provide evidence for the detection, characterization, and attribution decisions, and will collaborate with TA2 to integrate that evidence into results provided to an end user.

TA1 performers will be responsible for developing their own training data sets to build generalizable semantic DAC algorithms. TA1 performers will be expected to share their training data with all SemaFor performers, to support analysis, retraining, and deployment of SemaFor capabilities.

TA1 performers will need to demonstrate significant DAC performance capabilities on both news stories and social media in Phase 1 (see Figure 2 for performance goals). In Phase 2, TA1 performers will need to improve their evaluation performance and extend their techniques to new domains: technical information and collections of news articles. Technical information is highly relevant to the DoD and IC, but brings additional technical challenges as it involves highly specialized semantic domains. Technical information could include, for example, a news story from a nation state describing their latest missile capabilities. In Phase 3, TA1 performers will further improve their performance and prepare for possible transition.

Mandatory evaluations will include program-wide evaluations every eight months. TA1 performers must also participate in hackathons every six months as shown in Figure 3.

Hackathon participation may include software integration, evaluation dry runs, and challenge problem execution. Hackathons are expected to be one week in duration.

HR001119S0085 SEMAFOR 11

Strong TA1 proposals will describe:

Approaches to automatically reason about extraction failures in one or more modalities that might otherwise indicate spurious inconsistencies across modalities.

Approaches to align, ground, and reason about entities across multiple modalities, each of which might only have a portion of the overall narrative.

Algorithms for DAC that provide effective performance even with limited training data, and that are robust against domain mismatch.

DAC algorithms that could deal with real-world issues such as multiple cultures and contexts.

Techniques for quantitatively characterizing key aspects of falsified media, such as malicious intent, in ways that are both computationally accessible and operationally relevant.

TA2 Explanation and Integration

TA2 is responsible for developing algorithms that fuse scores across the multiple TA1 performers to provide a single fused detection score, which will be used to support analyst prioritization and review. Similar fusion will happen for attribution and characterization scores.

TA2 is also responsible for developing algorithms that automatically assemble and curate the evidence provided by the TA1 components into a summary explanation for an analyst. This work will support TA2’s development of algorithms that leverage scores and evidence from the TA1 performers to prioritize falsified media for human review. Such prioritization is critical for scaling up to real-world volumes of media.

TA2 proposals should describe in detail how their fusion approaches will combine information across multiple TA1 performers for each of the DAC tasks. In particular, the TA2 performer may be faced with a different number of TA1 performers responding to each probe; likely, there will not be large development data sets available to train fusion approaches. TA2 score fusion algorithms will be evaluated with the same metrics as the TA1 DAC algorithms. DARPA expects the TA2 performer will demonstrate significant performance or end user utility, or both, with their fusion approaches.

TA2 should expect that different prioritization schemes may be needed for different CONOPs or different users. TA2 proposals should describe how their prioritization algorithms will leverage DAC scores and evidence to identify and rank results that will be of most interest to targeted end users.

TA2 is also responsible for system integration and user interface development. TA2 will facilitate program design discussions and lead the design of system APIs that integrate TA1 and TA2. DARPA seeks a prototype SemaFor system that is highly scalable and that targets cloud-based deployment. Likely transition platforms include GovCloud and C2S, but may also include operational systems for DoD and Intelligence Community organizations. Compelling TA2

HR001119S0085 SEMAFOR 12

proposals will demonstrate a performer’s ability to transition SemaFor technologies to DoD and IC organizations at the Sensitive Compartmented Information level. Transition partners may also include commercial organizations, such as internet platforms. TA2 will collaborate closely with TA1 performers and will lead the design of SemaFor system APIs.

DARPA anticipates evaluating TA1 and TA2 algorithms in the context of the prototype SemaFor system. This evaluation process will help ensure that the algorithm assessments reflect the performance of a deployed SemaFor system. As such, TA2 will be expected to regularly receive, validate, and integrate components from TA1 performers into a prototype SemaFor system. A stable prototype system must be in place prior to each program evaluation.

TA2 will be responsible for supplying compute and data storage resources for SemaFor evaluations, hackathons, and demonstrations. See the schedule in Figure 3. DARPA seeks to enable a continuous integration, continuous deployment, continuous evaluation process on the program, while also minimizing the cost of compute.

TA2 will be responsible for hosting and leading program hackathons. The purpose of the hackathons is to support the development and integration of SemaFor components and to execute challenge problems from TA4. Hackathons will include performers from all TAs. Hosting hackathons will involve, at a minimum, providing work and meeting space, and network capacity to support 50 people for one week for each hackathon.

TA2 will be responsible for supporting integration exercises with transition partners. Proposers should budget for two such exercises in the Washington, D.C. area, to include software installation, training, and support (some onsite; some remote) for a three-month test and evaluation period.

Strong TA2 proposals will describe:

Techniques for fusing DAC scores across multiple TA1 performers each with disparate approaches.

Approaches for reconciling evidence across multiple TA1 performers with disparate forms of evidence, and presenting a unified evidence summary and explanation to end users.

Methods for automatically customizing media prioritization schemes to different end users or different classes of end users.

Technical approaches to enabling parallel TA1 development and system integration while simultaneously minimizing dependencies and integration effort.

A strategy for supporting a rolling, continuous evaluation process that leverages the prototype SemaFor system and a continuous integration, continuous deployment process while keeping compute costs in check.

HR001119S0085 SEMAFOR 13

A strategy for storage of training and evaluation data as well as initialization values, hyper-parameters, training and evaluation process scripts, documentation, validation tests, knowledge-bases, and any other materials required for the Government to reproduce or retrain any algorithmic component. Data will need to be stored securely and also be able to be compartmentalized to ensure that evaluation data is kept separate from training data.

An approach for proactively engaging with potential transition customers to enable early transition of SemaFor capabilities.

Evidence of previously successful transition of DARPA capabilities to operational use in the DoD and/or IC.

TA3 Evaluation

TA3 will collaborate closely with DARPA, TA1, TA2, TA4, and potential transition partners in the design and execution of the program evaluations. The purpose of the evaluations is to understand how well SemaFor capabilities might meet the needs of potential transition partners, such as DoD, IC, and commercial organizations, and to understand the program’s progress against its scientific goals. SemaFor evaluations will characterize all elements of the prototype SemaFor system. TA3 will be expected to facilitate and lead program discussions about evaluation design, including data sets, evaluation processes, evaluation schedule, metrics, and transition partner use cases.

Figure 2: Program goals for DAC

In addition to the required metrics shown in Figure 2, TA3 proposals should describe metrics that are relevant for assessing the performance of a prototype SemaFor system. Proposers should describe how the additional metrics will help advance the understanding of a prototype SemaFor system in operational use or enhance scientific understanding of SemaFor algorithms.

In particular, DARPA is interested in metrics that augment the DAC metrics in Figure 2 and

HR001119S0085 SEMAFOR 14

inform the performance of TA2s fusion, explanation, and prioritization components. TA3 will be responsible for implementing the program metrics as well as any additional proposed metrics.

The Government will provide the image and video scoring code developed on the DARPA Media Forensics (MediFor) program.

DARPA is also interested in understanding where human capabilities might be best augmented by automated algorithms. Therefore, TA3 performers will design experiments to establish baseline human performance for DAC of falsified media. TA3 proposers should present a design for human subject research that would establish relevant human baseline performance. DARPA expects that assessing typical adult performance would provide a useful human baseline, but TA3 proposers should justify their own experimental design in terms of experiments, metrics, and participants. TA3 performers should provide a plan for how they will quickly obtain any necessary Institutional Review Board (IRB) approvals for both algorithm evaluations and human baseline experiments such that planned program evaluations can be conducted as scheduled.

For planning purposes, human baseline experiments could be conducted towards the end of Phase 1.

In support of evaluations, TA3 will also be responsible for media generation and curation.

DARPA will provide the existing image and video media developed on the DARPA MediFor program. Other sources of media such as Creative Commons licensed works and open source resources should be investigated and used when possible. Any additional needs for media will be met through media created for the SemaFor program by personnel employed by the TA3 performer. TA3 will provide sample evaluation data to TA1 and TA2 performers to enable algorithm testing, integration, and dry-run evaluation in addition to the media reserved for the program evaluations.

Based on experience from previous DARPA programs, the following dataset sizes are considered the minimum necessary for a sufficient evaluation. For background world media, the TA3 performer is expected to collect 250,000 news articles and 250,000 social media posts during Phase 1. The TA3 performers will falsify approximately 2,500 news articles and 2,500 social media posts in the first phase of the program. For Phase 2, the TA3 performer is expected to collect an additional 25,000 technical information articles and falsify approximately 2,000 technical information articles. Similarly in Phase 2, the TA3 performer is expected to collect an additional 25,000 news collections (with a collection being two or more articles) and to falsify approximately 2,000 news collections. In Phase 3, the TA3 performer is expected to collect an additional 25,000 news collections and 10,000 technical information collections. However, proposals must provide justification for the data plan offered, and should describe in detail how the data plan would support the evaluation proposed and the program goals, including scaling to address new manipulation threats.

News articles should span a range of local, national, and international events with a particular focus on stories where falsification could have significant real-world impact. Collected news articles should have as much context as possible, in particular the URL, source organization, author, date, and location (if provided). Social media assets should also focus on local, national, and international events where falsification could have a significant real-world impact.

Technical information should describe technical capabilities of interest to the defense and

HR001119S0085 SEMAFOR 15

intelligence communities. All collected or falsified assets should be multi-modal, containing at least two media modalities. TA3 proposers should describe the content their collection and falsification strategies will focus on, and how that content will inform the evaluation design.

TA3 proposals should include a detailed discussion of the experimental design planned to test the TA1 and TA2 components. The discussion should include the types of probes that will be selected and how the selection process relates to expected real-world media falsification.

Proposals will need to put forth a method for handling the potential combinatorial complexity of evaluating performers on multiple media and falsification types, in cross-modality media groupings of various compositions. TA3 will conduct evaluations every 8 months with an initial dry run 6 months into the program. Strong proposals will describe a process for implementing rolling evaluations, including a performance leaderboard updated as algorithms as submitted and automatically evaluated.

The target milestones and metrics identified above have been established to assess technical progress over the course of the program. These targets represent the expected pace of technology development. The targets are not “go/no-go” criteria, and it is not DARPA’s intention to use these targets as a basis for down-selects.

TA3 will organize and host Principal Investigator (PI) meetings every 6 months for all of the performers, DARPA, and invited guests from other government agencies and industry. PI meetings will last two or three days each. Meeting locations may vary based on the locations of the performers. Meeting attendance is estimated to be approximately 175 people.

Strong TA3 proposals will describe:

A detailed plan for obtaining and curating data that is sufficient in volume, highly relevant to the problem domain, and can be released to the broader research community during the course of the program. The plan should include estimates for how many news articles, social media probes, and technical information articles will be needed to support evaluations in each phase of evaluation.

How the evaluation design will identify, manage, and decouple latent variables that might be unintentionally correlated across evaluation probes.

The evaluation team’s approach to having strong subject matter expertise in the detection, attribution, characterization, explanation, and prioritization of falsified multi-modal media.

How the evaluation design and roadmap will provide both a comprehensive understanding of the program’s scientific progress and answer key performance questions for potential transition partners.

Strategies for designing, organizing, and executing complex evaluation processes across a large distributed team while maintaining performer buy-in and evaluation integrity.

HR001119S0085 SEMAFOR 16

Approaches for streamlining the human subjects research and IRB process related to evaluation.

A process for data safeguarding that will prevent data loss or unintended leakage of data outside of the program.

TA4 Challenge Curation

TA4 will curate state-of-the-art (SOTA) challenges drawn from the public domain to ensure that the SemaFor program addresses relevant threat scenarios. TA4 will also develop threat models, based on current and anticipated technology, to help ensure that SemaFor defenses will be highly relevant for the foreseeable future. TA4 will include multiple challenge problem curation teams who will collaborate to maximize coverage of the challenge space and threat models.

TA4 will regularly deliver challenges and updated threat models to the TA3 evaluation team and

DARPA.

The challenges and threat models developed by TA4 will be used to drive the hackathons run by TA2 with support from TA3. TA4 performers will deliver SOTA challenges (and supporting threat models if relevant) to DARPA and to TA3 starting at month 4 of the program and then at least every 6 months following for the duration of the program. Support for the hackathons will involve working with TA3 to curate additional generated or manipulated media for the challenge problems. If existing media is not sufficient to support the challenge, TA4 will work with TA3 to generate new media to support the challenge.

During the hackathons, TA4 will be expected to lead the use of challenge problems. TA4 will coordinate in advance of the hackathon to ensure that the upcoming challenge is understood by TA1 and TA2. TA4 will be on-site during hackathons to answer questions from TA1 and TA2 participants, and also to work with TA3 to evaluate progress on the challenge. TA4 will coordinate with TA3 to include useful challenges into program-wide evaluations. Hackathon challenges will provide a way to flight-test a problem before program-wide evaluation.

TA4 threat models are intended to provide insight to the program as to where and how automated DAC technologies could be most effective. TA4 threat models should inform the assessment of SemaFor technologies against current and future threats. Threat models could include information such as detailed insight into the supply chain for a manipulation to inform where and how defenses might be most effective. Threat models could provide insight into the strengths and weaknesses of human abilities to detect, attribute, and characterize falsified media. Threat models should anticipate coming advances in manipulation technology and how those advances might impact current and future defenses. Threat models should inform the program about where additional burdens can be placed on a manipulator to increase our ability to detect, attribute, or characterize falsified media or to degrade an adversary’s ability to generate falsified media at scale.

Particularly strong TA4 proposals will bring demonstrated knowledge of media threats to national security and stabilization efforts. Experience in identifying threats, creating threat models (both after the fact and speculative), and collecting or generating media are all desired.

HR001119S0085 SEMAFOR 17

Strong TA4 proposals will describe:

Detailed evidence of the proposer’s ability to bring state-of-the-art falsification challenges in one or more modalities to the program.

Threat models that provide actionable insights into how DAC algorithms and the SemaFor system should be designed to put significant burdens on potential manipulators.

Schedule/Milestones

Figure 3 SemaFor Program Schedule

Program hackathons will enable performers to collaborate directly on system design, software integration, challenge problems, and evaluation dry-runs. Performers should plan that hackathons are approximately a work week in duration, starting Monday afternoon and completing Friday morning. Hackathons will be organized by the TA2 team, and an agenda for the meeting will be provided in advance of the event. Hackathons will be used as both software integration and design exploration, and as challenge problems or evaluation dry-runs.

Deliverables and TA Performer Interactions

TA1 performers will deliver algorithms that detect, characterize, and attribute falsified multi-modal media. Algorithms will ideally be under constant development and improvement, and mandatory drops to TA2 for system integration will be needed ahead of each evaluation. Each TA1 performer will work with the TA2 performer to integrate their algorithms, knowledge, resources, and other results into the TA2 performer's workflow.

The TA2 performer will work with the TA1 performers to integrate algorithms, knowledge, and resources from TA1 performers into the TA2 system. TA2 will deliver algorithms that:

Fuse the DAC scores provided by TA1 performers.

HR001119S0085 SEMAFOR 18

Automatically assemble and curate the evidence provided by the TA1 components into a summary explanation for an analyst.

Support prioritization of media for analyst review.

TA2 will also deliver periodic proof-of-concept systems that integrate multiple TA1 components into a SemaFor system targeting scalable cloud deployment. TA2 will be expected to provide a demonstration to the government in each program phase of the progressing capabilities of the SemaFor system. SemaFor evaluations will be conducted by evaluating TA1 and TA2 algorithms integrated into the SemaFor system. In support of the evaluations, TA2 will provide the integrated system and data processing capabilities necessary to carry out evaluations.

Integrated systems are required at least two weeks before scheduled evaluations.

TA3 will design, organize, plan, and conduct the SemaFor evaluations and results analysis. While the evaluations will be conducted at the TA2 performer’s location, they will be under the control and supervision of the TA3 performer. TA3 deliverables include program metrics, evaluation protocols, and a library of multi-modal media assets for development and test purposes. Media may be collected or created. TA3 must prepare and deliver an evaluation design and schedule no less than 60 days prior to each evaluation. TA3 must prepare and deliver an analysis report of the results to DARPA no later than 20 business days after the conclusion of each evaluation. TA3 will be responsible for the ACA process across the program.

TA4 will deliver challenge problems and threat models in support of hackathons and program evaluations. TA4 will develop multi-modal media, in collaboration with TA3, to support the challenge problems and threat model development. TA4 will provide TA3 with expertise for the development of the evaluation protocols and test items. TA4 is expected to work with TAs 1 and 2 to improve their technology and understanding of the problem domain.

All performers (TA1, TA2, TA3, and TA4) shall be required to provide the following deliverables via DARPA’s Technical-Financial Information Management System (TFIMS) database:

Technical briefings and reports - Kickoff presentations with changes or updates shall be submitted within 1 month of the program kickoff meeting. Presentation materials from PM site visits shall be submitted within 1 month of the review.

Quarterly progress reports - A quarterly progress report describing progress made, resources expended, and issues requiring the attention of the Government team shall be provided within 15 days of the end of each fiscal quarter.

Monthly financial reporting.

Final report - The final report shall concisely summarize the effort.

In addition to reports, TA1 and TA2 performers shall be required to provide the following deliverables to DARPA:

Intermediate and final versions of software libraries, source code, data, all build scripts, HR001119S0085 SEMAFOR 19 and sufficient documentation to enable the government to compile the source code into executable binaries as well as re-create any containerized software. Intermediate versions are due 1 month after each evaluation.

Training data, initialization values, hyper-parameters, training process scripts, documentation, validation tests, knowledge-bases, and any other materials required for the Government to reproduce or retrain any algorithmic component.

Implementation documentation - Documentation shall be provided one month after each code drop documenting any algorithms, source code, hardware descriptions, language specifications, system diagrams, part numbers, and other data necessary to replicate and test the designs.

Application user manual and training material.

In addition to reports, TA3 and TA4 performers shall be required to provide data, falsified media, and annotations every 3 months as detailed in the TA3 and TA4 descriptions.

to TA1 to TA2 to TA3 to TA4 to Program TA1 provides Input into system

API

specification/design

DAC algorithm containers implementing API and documentation

DAC scores and evidence via API

Support for DAC container integration

TA1 algorithm insight to support fusion, explanation, and prioritization components

Calibrated scores

Suggested datasets and evaluation scenarios

Support for designing evaluations

Feedback on challenges and Hackathons

Development & training data

Participation at hackathons and PI meetings

Develop and provide insight into DAC algorithms

DAC algorithm containers implementing API and documentation

TA2 provides System API specifications designed with TA1 input

Integration of

TA1

components into SemaFor system

Compute resources for evaluation of TA1 algorithms

Design input for score calibration processLead

Support for designing evaluations

Compute for evaluation scoring code

Feedback on challenges and Hackathons

SemaFor system design and APIs

SemaFor system integration and U/I development

Provide compute resources for evaluations, hackathons, and demonstrations

Transition support Support integration exercises with transition partners

Hosting and leading hackathons

HR001119S0085 SEMAFOR 20

to TA1 to TA2 to TA3 to TA4 to Program design of system APIs

Receive validate and integrate TA1 components into SemaFor system

Participation at hackathons and PI meetings

Develop and provide insight into score fusion, explanation, and prioritization algorithms

SemaFor system demonstrations in each program phase

Develop algorithms to assemble and curate evidence; provide unified evidence summary and explanation

Facilitate program design discussions

Provide a stable prototype system prior to each evaluation

TA3 provides Sample development and evaluation data

Sample development and evaluation data

Input to challenges and Hackathons

Media generation and curation

Facilitate and lead program discussions about evaluation designs (datasets, processes, schedule, metrics, transition partner use cases)

Define and implement metrics

Design and conduct experiments to establish baseline human performance

Evaluation scoring software

Evaluation results analysis

Organize and host PI meeting

Oversight of PI meetings

Conduct evaluations every 8 months

TA4 provides Coordinate in advance of and during Hackathons to ensure challenge understanding

Coordinate in advance of and during Hackathons to ensure challenge understanding

Support for incorporating challenge problems into evaluations

Regularly deliver

State of the art falsification techniques

Curate SOTA challenges from public domain

HR001119S0085 SEMAFOR 21

to TA1 to TA2 to TA3 to TA4 to Program challenges and updated threat models

Work with TA3 to curate additional generated or manipulated data for challenge problems

Work with TA3 to evaluate progress on challenge

Develop threat models

Provide insight as to whether/how DAC technologies could be most effective

Challenge problem design

Lead challenge problem execution at hackathons

Participation at hackathons and PI meetings

Figure 4 Technical Area interactions

Government-furnished Property/Equipment/Information

The Government will provide the DARPA MediFor system APIs, algorithms, and software for potential inclusion in TA1 efforts and as a starting point for the SemaFor system for TA2 efforts.

The Government will provide media (images and video) developed on the MediFor program to all performers.

Intellectual Property

The program will emphasize creating and leveraging open source technology and architecture.

Intellectual property rights asserted by proposers are strongly encouraged to be aligned with open source regimes. See Section VI.B.1 for more details on intellectual property. Any proposer claiming the use of proprietary technology or choosing to explicitly exclude their technology from the open source regime will need to provide adequate justification in their proposal.

A key goal of the program is to establish an open, standards-based, multi-source, plug-and-play architecture that allows for interoperability and integration. This includes the ability to easily add, remove, substitute, and modify software and hardware components. This will facilitate rapid innovation by providing a base for future users or developers of program technologies and deliverables. Therefore, it is desired that all noncommercial software (including source code), software documentation, hardware designs and documentation, and technical data generated by the program be provided as open source deliverables to the Government, as lesser rights may adversely impact the lifecycle costs of affected items, components, or processes.

HR001119S0085 SEMAFOR 22

II. Award Information

A. Awards

DARPA anticipates multiple awards for TA1 and TA4, as well as single awards for TA2 and TA3.

The level of funding for individual awards made under this solicitation has not been predetermined and will depend on the quality of the proposals received and the availability of funds. Awards will be made to proposers whose proposals are determined to be the most advantageous to the Government, all factors considered, including the potential contributions of the proposed work, overall funding strategy, and availability of funding. See Section V for further information.

The Government reserves the right to:

select for negotiation all, some, one, or none of the proposals received in response to this solicitation;

make awards without discussions with proposers;

conduct discussions with proposers if it is later determined to be necessary;

segregate portions of resulting awards into pre-priced options;

accept proposals in their entirety or to select only portions of proposals for award;

fund proposals in increments and/or with options for continued work at the end of one or more phases;

request additional documentation once the award instrument has been determined

(e.g., representations and certifications); and remove proposers from award consideration should the parties fail to reach agreement on award terms within a reasonable time or the proposer fails to provide requested additional information in a timely manner.

Proposals selected for award negotiation may result in a procurement contract, grant, cooperative agreement, or Other Transaction (OT) depending upon the nature of the work proposed, the required degree of interaction between parties, and other factors.

In all cases, the Government contracting officer shall have sole discretion to select award instrument type, regardless of instrument type proposed, and to negotiate all instrument terms and conditions with selectees. DARPA will apply publication or other restrictions, as necessary, if it determines that the research resulting from the proposed effort will present a high likelihood of disclosing performance characteristics of military systems or manufacturing technologies that are unique and critical to defense. Any award resulting from such a determination will include a requirement for DARPA permission before publishing any information or results on the program. For more information on publication restrictions, see the section below on Fundamental Research.

HR001119S0085 SEMAFOR 23

B. Fundamental Research

It is DoD policy that the publication of products of fundamental research will remain unrestricted to the maximum extent possible. National Security Decision Directive (NSDD) 189 defines fundamental research as follows:

‘Fundamental research’ means basic and applied research in science and engineering, the results of which ordinarily are published and shared broadly within the scientific community, as distinguished from proprietary research and from industrial development, design, production, and product utilization, the results of which ordinarily are restricted for proprietary or national security reasons.

As of the date of publication of this BAA, the Government expects that program goals as described herein may be met by proposers intending to perform fundamental research and does not anticipate applying publication restrictions of any kind to individual awards for fundamental research that may result from this BAA.