Participating Lab: Air Force Research Laboratory Information Directorate (AFRL RI)

In the era of AI and Zero-Trust, the principles of communication anonymization and covertness are more critical than ever. As analysts leverage powerful, often third-party, AI tools for sensitive tasks, the communication patterns themselves, accessing specific AI services, query frequencies, data volumes, become a significant source of metadata leakage. A true zero-trust model cannot implicitly trust the network to hide user intent. Therefore, robust protocols are needed to anonymize these interactions and provide a foundational layer of security upon which privacy-preserving computation can operate.
This research topic seeks novel protocols that integrate AI to achieve unprecedented levels of communication security in zero-trust environments. We are particularly interested in solutions that can counter AI-powered adversaries who use machine learning for traffic analysis and deanonymization.
Specific areas of interest include, but are not limited to:
•AI-Enhanced Anonymity Networks: Designing and evaluating novel architectures that use AI to dynamically reroute traffic, predict and counter sophisticated timing attacks, and optimize for low-latency, high-throughput performance.
•Generative AI for Steganography and Covert Channels: Leveraging generative models (e.g., GANs, Diffusion Models) to create highly realistic cover data (images, text, network traffic) for steganographic channels, making them statistically indistinguishable from benign data streams.
•Deniable Communication in a Zero-Trust Framework: Creating protocols that allow parties to plausibly deny communication, with a focus on resisting AI-driven forensic analysis of network artifacts.
•Explainable AI for Anonymity Verification: Using Explainable AI to formally verify the anonymity and covertness properties of protocols, providing a trustworthy assurance level rather than relying on statistical arguments alone.
•Metadata Obfuscation against AI-Powered Traffic Analysis: Developing innovative methods to outsmart ML-based classifiers by altering metadata signatures (timing, volume, patterns) to prevent the leakage of sensitive operational intent.

Zero-trust is a security model that rigorously verifies every user and device before granting access to resources. Within cloud environments, this principle dictates that no entity, whether internal or external, is trusted by default. While offering robust security, practical implementation often relies on complex technologies like public-key infrastructure and zero-knowledge proofs. This challenge is magnified when Artificial Intelligence (AI) is used to automate security decisions, as the “”black-box”” nature of many models creates an implicit trust that contradicts the core “”never trust, always verify”” tenet.
This research topic seeks novel approaches to: 1) enabling warfighters to securely outsource private data and computation to untrusted clouds with mission assurance and verifiable correctness; 2) improving the resilience of mission-critical applications by distributing them across multiple heterogeneous cloud providers to prevent single points of failure; 3) optimizing the trade-off between strict zero-trust security and performance for time-sensitive missions; and 4) ensuring that any AI used to enforce the security framework is itself explainable, verifiable, and trustworthy. Research topics of interest include, but are not limited to:
•Explainable and Verifiable AI for Zero-Trust Policy Enforcement: Research into AI/ML models that not only automate access control and threat detection but also provide verifiable proof of their reasoning, ensuring the security framework itself is auditable and does not rely on “”black-box”” decisions.
•Decentralized identity and access control mechanisms and protocols, including those that support anonymity.
•Novel cryptographic primitives and protocols with application to zero-trust and AI-driven security paradigms.
•Design of cross-cloud, privacy-aware protocols that integrate with emerging zero-trust security mechanisms and provide verifiable attestations of AI behavior.
•End-to-end data protection, concurrency, and consistency for multi-user, multi-cloud environments.
•Access patterns and volume leakage prevention for oblivious data stores under malicious security.
•Verification and authentication schemes for AI-driven query evaluation over encrypted and unencrypted data.

The modern information landscape presents a dual challenge for the Air Force: (1) the proliferation of sophisticated, multimodal AI-generated media (deepfakes) poses a significant threat to information integrity, and (2) the sheer volume of collected video and sensor data makes finding specific, mission-relevant events nearly impossible without exhaustive manual review. This research program addresses both challenges through a unified framework, leveraging Neuro-Symbolic AI to create systems that can not only forensically analyze multimodal content but also perform high-fidelity, prompt-based retrieval from massive video archives.
This topic seeks novel approaches that move beyond surface-level artifact detection. Instead, we aim to build systems that combine the pattern-recognition strengths of neural networks with the contextual reasoning of symbolic AI. This will enable a deeper understanding of relationships, objects, and actions within and across different media types (video, images, audio, text). The goal is to develop capabilities for both defending against AI-generated disinformation and for turning petabytes of raw footage into actionable, searchable intelligence.
Research areas of primary interest include:
• Advanced Multimodal Forensics & Red Teaming: This thrust focuses on the detection of AI-generated content that is explicitly designed to be undetectable. This includes developing novel generative techniques (e.g., via cross-modal diffusion models or GANs) to create stealthy synthetic media, which will then be used to red team and future-proof our detection systems. The goal is to build explainable forensic models that identify inconsistencies not just within a single modality, but across video, audio, and text domains.
• Neuro-Symbolic Video Analysis and Retrieval: This addresses the data overload problem by developing AI that can understand and respond to complex, natural-language prompts. The goal is to enable an analyst to retrieve a specific, short clip from hours of footage by asking a question like, “Show me all instances where a blue vehicle stops and a person exits carrying a package.” This requires the AI to build a structured, symbolic representation of the video’s content, enabling it to reason about objects, actions, and their temporal relationships without relying on simple keyword matching.
• Explainable and Verifiable AI: A critical component for both forensics and retrieval is trust. Any system developed must be able to explain why it flagged an image or video as synthetic or how it interpreted a prompt to find a specific result. This thrust will focus on creating verifiable models whose reasoning can be audited and understood by a human operator, ensuring that the AI is a trustworthy partner in the analytical process.

This research topic seeks novel methodologies for conducting comprehensive and secure analysis across the vast landscape of Publicly Available Information (PAI). PAI is inherently multilingual and takes diverse forms, including online publications, news, social media, and AI-generated content. The central challenge is twofold: first, to enable analysts to perform PAI retrieval while rigorously protecting their identity and intent from being tracked or linked back to them. Second, as analysts increasingly rely on advanced AI models (e.g., Large Language Models) for analysis, a new challenge emerges: protecting the query content and analytical intent from the AI service providers themselves. This requires a new generation of privacy-preserving technologies that can operate on untrusted platforms.
We are particularly interested in novel technologies and approaches that address:
•Privacy-Preserving ML for PAI Analysis: Advanced cryptographic techniques for secure ML training and inference on PAI. This includes, but is not limited to, Homomorphic Encryption (HE) for computation on encrypted data, Secure Multi-Party Computation (MPC), and hybrid cryptographic protocols that combine the strengths of both.
•Privacy-Preserving Multimodal LLMs: Methods to query and interact with Large Language Models (both “”black-box”” APIs and “”white-box”” local models) without revealing the sensitive content or context of the user’s query. This must cover multimodal inputs, including text, images, and other data types.
•Privacy-Preserving Semantic Search: Protocols and algorithms for conducting secure vector-based similarity searches. This is critical for RAG and other AI applications, enabling an analyst to find relevant information without exposing their query vector to the database service.
•Confidential and Scalable PAI Collection: Methods for confidential PAI gathering that do not rely on trusted third parties, modify source data, or require collaboration with website operators.
•Robust Attribution Management and Anonymization: Advanced approaches for managing attribution and ensuring comprehensive web browsing anonymization to protect the analyst at the network level.
•Private Information Retrieval (PIR): Innovative PIR protocols to facilitate secure and private data access from public and untrusted databases.

To ensure information dominance in future conflicts, U.S. Air Force ISR capabilities must be able to operate effectively within highly contested and electronically hostile environments. This research program addresses two critical gaps preventing this dominance: (1) the inability to seamlessly fuse near-field (NF) and far-field (FF) Synthetic Aperture Radar (SAR) data for a complete multi-scale picture, and (2) the growing vulnerability of SAR systems to sophisticated electronic warfare and adversarial attacks. Current systems lack a unified framework to bridge these scales, while facing emerging threats that can corrupt raw sensor data to create deceptive “”ghost”” targets or conceal real ones. This research topic will develop a unified, resilient framework that integrates solutions to both challenges, enabling high-fidelity characterization and targeting even when facing active opposition.
A primary technical hurdle is reconciling the divergent physics of NF (spherical wavefront) and FF (planar wavefront) sensing, especially when data is degraded by active electronic attack such as barrage jamming or GPS-denial. This research will address this by developing a hybrid imaging chain with integrated logic to adapt to interference through robust beamforming and adaptive filtering. A further focus will be on the novel threat of waveform-domain adversarial attacks. The research will establish a taxonomy of practical, physics-based attacks and create a high-fidelity simulation testbed for generating and analyzing them. Based on this, new reconstruction algorithms using non-linear techniques, Bayesian inference, or deep unfolding will be designed to be inherently robust to these manipulations. These algorithms will be part of a larger effort to develop “”threat-aware”” sensor management, where the system can autonomously detect, flag, and even adjust its collection strategy to nullify detected adversarial injections.
Upon completion, this research will deliver a battle-ready, hardened multi-scale SAR architecture. The key outcomes will be a processing chain that is demonstrably resilient to both jamming and adversarial manipulation, an adversarial SAR testbed for red-teaming future systems, and a library of validated counter-adversarial algorithms. The ultimate impact will be a demonstrated leap in capability, providing commanders with reliable, trustworthy, and high-fidelity intelligence for targeting and decision-making, ensuring information superiority in the most challenging operational environments.