Matthew Green

Soham Hans, Volkan Ustun, Benjamin Nye et al.

Achieving expert-level performance in simulation-based training relies on the creation of complex, adaptable scenarios, a traditionally laborious and resource intensive process. Although prior research explored scenario generation for military training, pre-LLM AI tools struggled to generate sufficiently complex or adaptable scenarios. This paper introduces a multi-agent, multi-modal reasoning framework that leverages Large Language Models (LLMs) to generate critical training artifacts, such as Operations Orders (OPORDs). We structure our framework by decomposing scenario generation into a hierarchy of subproblems, and for each one, defining the role of the AI tool: (1) generating options for a human author to select from, (2) producing a candidate product for human approval or modification, or (3) generating textual artifacts fully automatically. Our framework employs specialized LLM-based agents to address distinct subproblems. Each agent receives input from preceding subproblem agents, integrating both text-based scenario details and visual information (e.g., map features, unit positions and applies specialized reasoning to produce appropriate outputs. Subsequent agents process these outputs sequentially, preserving logical consistency and ensuring accurate document generation. This multi-agent strategy overcomes the limitations of basic prompting or single-agent approaches when tackling such highly complex tasks. We validate our framework through a proof-of-concept that generates the scheme of maneuver and movement section of an OPORD while estimating map positions and movements as a precursor demonstrating its feasibility and accuracy. Our results demonstrate the potential of LLM-driven multi-agent systems to generate coherent, nuanced documents and adapt dynamically to changing conditions, advancing automation in scenario generation for military training.

Brennon Brimhall, Philip Mathew, Neil Fendley et al.

Machine unlearning methods take a model trained on a dataset and a forget set, then attempt to produce a model as if it had only been trained on the examples not in the forget set. We empirically show that an adversary is able to distinguish between a mirror model (a control model produced by retraining without the data to forget) and a model produced by an unlearning method across representative unlearning methods from the literature. We build distinguishing algorithms based on evaluation scores in the literature (i.e. membership inference scores) and Kullback-Leibler divergence. We propose a strong formal definition for machine unlearning called computational unlearning. Computational unlearning is defined as the inability for an adversary to distinguish between a mirror model and a model produced by an unlearning method. If the adversary cannot guess better than random (except with negligible probability), then we say that an unlearning method achieves computational unlearning. Our computational unlearning definition provides theoretical structure to prove unlearning feasibility results. For example, our computational unlearning definition immediately implies that there are no deterministic computational unlearning methods for entropic learning algorithms. We also explore the relationship between differential privacy (DP)-based unlearning methods and computational unlearning, showing that DP-based approaches can satisfy computational unlearning at the cost of an extreme utility collapse. These results demonstrate that current methodology in the literature fundamentally falls short of achieving computational unlearning. We conclude by identifying several open questions for future work.

Maximilian Zinkus, Tushar M. Jois, Matthew Green

Mobile devices have become an indispensable component of modern life. Their high storage capacity gives these devices the capability to store vast amounts of sensitive personal data, which makes them a high-value target: these devices are routinely stolen by criminals for data theft, and are increasingly viewed by law enforcement agencies as a valuable source of forensic data. Over the past several years, providers have deployed a number of advanced cryptographic features intended to protect data on mobile devices, even in the strong setting where an attacker has physical access to a device. Many of these techniques draw from the research literature, but have been adapted to this entirely new problem setting. This involves a number of novel challenges, which are incompletely addressed in the literature. In this work, we outline those challenges, and systematize the known approaches to securing user data against extraction attacks. Our work proposes a methodology that researchers can use to analyze cryptographic data confidentiality for mobile devices. We evaluate the existing literature for securing devices against data extraction adversaries with powerful capabilities including access to devices and to the cloud services they rely on. We then analyze existing mobile device confidentiality measures to identify research areas that have not received proper attention from the community and represent opportunities for future research.

Maximilian Zinkus, Tushar M. Jois, Matthew Green

In this work we present definitive evidence, analysis, and (where needed) speculation to answer the questions, (1) Which concrete security measures in mobile devices meaningfully prevent unauthorized access to user data? (2) In what ways are modern mobile devices accessed by unauthorized parties? (3) How can we improve modern mobile devices to prevent unauthorized access? We examine the two major platforms in the mobile space, iOS and Android, and for each we provide a thorough investigation of existing and historical security features, evidence-based discussion of known security bypass techniques, and concrete recommendations for remediation. We then aggregate and analyze public records, documentation, articles, and blog postings to categorize and discuss unauthorized bypass of security features by hackers and law enforcement alike. We provide in-depth analysis of the data potentially accessed via law enforcement methodologies from both mobile devices and associated cloud services. Our fact-gathering and analysis allow us to make a number of recommendations for improving data security on these devices. The mitigations we propose can be largely summarized as increasing coverage of sensitive data via strong encryption, but we detail various challenges and approaches towards this goal and others. It is our hope that this work stimulates mobile device development and research towards security and privacy, provides a unique reference of information, and acts as an evidence-based argument for the importance of reliable encryption to privacy, which we believe is both a human right and integral to a functioning democracy.

Michael Specter, Sunoo Park, Matthew Green

Email breaches are commonplace, and they expose a wealth of personal, business, and political data that may have devastating consequences. The current email system allows any attacker who gains access to your email to prove the authenticity of the stolen messages to third parties -- a property arising from a necessary anti-spam / anti-spoofing protocol called DKIM. This exacerbates the problem of email breaches by greatly increasing the potential for attackers to damage the users' reputation, blackmail them, or sell the stolen information to third parties. In this paper, we introduce "non-attributable email", which guarantees that a wide class of adversaries are unable to convince any third party of the authenticity of stolen emails. We formally define non-attributability, and present two practical system proposals -- KeyForge and TimeForge -- that provably achieve non-attributability while maintaining the important protection against spam and spoofing that is currently provided by DKIM. Moreover, we implement KeyForge and demonstrate that that scheme is practical, achieving competitive verification and signing speed while also requiring 42% less bandwidth per email than RSA2048.

Katherine Q. Ye, Matthew Green, Naphat Sanguansin et al.

We have formalized the functional specification of HMAC-DRBG (NIST 800-90A), and we have proved its cryptographic security--that its output is pseudorandom--using a hybrid game-based proof. We have also proved that the mbedTLS implementation (C program) correctly implements this functional specification. That proof composes with an existing C compiler correctness proof to guarantee, end-to-end, that the machine language program gives strong pseudorandomness. All proofs (hybrid games, C program verification, compiler, and their composition) are machine-checked in the Coq proof assistant. Our proofs are modular: the hybrid game proof holds on any implementation of HMAC-DRBG that satisfies our functional specification. Therefore, our functional specification can serve as a high-assurance reference.