Fujimoto, Ted C. | Search NASA

DOE OSTI2024

Assurance of Reasoning Enabled Systems (ARES)

ARES was in part motivated by the determination of President’s Council of Advisors on Science and Technology (PCAST) on May 13th, 2023 that published a set of inquiries: In an era in which convincing images, audio, and text can be generated with ease on a massive scale, how can we ensure reliable access to verifiable, trustworthy information? How can we be certain that a particular piece of media is genuinely from the claimed source? What technologies, policies, and infrastructure can be developed to detect and counter AI-generated disinformation? In an effort to automatically analyze and patch/optimize code the work in this report describes various neural Machine Learning (ML) analysis engine implementations to assist in situations where source code is deficient or completely lacking to decompile (lift) binary code to ’C’. The goal is to gradually reduce human intervention. To this end, two Large Language Model (LLM) variants (Code LLama 2, LLama 3.1 and Starcoder1, Starcoder 2) where finetuned with ’before/after’ code pairs on the OpenBLAS library. LLama trained on the lowering process, Starcoder trained on the lifting process with National Security Agency’s (NSA) open-source Ghidra decompiler assist. The inferencing test results indicate correctness for only very short sequences for Starcoder 2. Moving forward, the experiments conclude with a set of recommendations of required resources and technologies

97 MATHEMATICS AND COMPUTING↗

DOE OSTI2021

The Effect of Antagonistic Behavior in Reinforcement Learning

The significant achievements of deep reinforcement learning (RL) have motivated researchers to also investigate its shortcomings. Such work has shown that typical methods in deep RL tend to produce brittle policies that overfit to the training environment. In this paper, we introduce the notion of purely antagonistic behavior in value-based agents, where the objective is not to maximize reward but to minimize the victim’s value over time. This notion is motivated by the scenario in which an antagonistic human architect, without access to the environment’s reward function, wants to build an RL agent that can impede another well-trained RL victim agent. First, we formalize a notion of antagonistic behavior in RL. Then, we provide experiments that show how a purely antagonistic agent performs compared to a well-trained victim that learns directly from the game’s rewards. Our results suggest that if one’s goal is to find vulnerabilities in well-trained agents, direct access to the environment’s rewards is not necessary, and antagonistic behavior can be measured independently from environment wins and losses.

Fujimoto, Ted C.↗