Trojans in Artificial Intelligence

2026

Open publication workspace · Sign in to read the full PDF.

AI-generated summary

1) This report details the findings of the TrojAI program, a multi-year initiative by IARPA to understand and mitigate the threat of AI Trojans, which are malicious, hidden backdoors embedded in AI models.
2) * The report covers the program's history, research focus, and related work in AI security.
* It details methodologies for detecting and mitigating AI Trojans, including weight analysis and trigger inversion techniques.
* The report also includes a thorough data analysis of performer submissions, examining detector performance, sensitivity, and the nature of "natural" Trojans.
3) AI Trojans, backdoor attacks, detection, mitigation, machine learning security

Check the original publication for accuracy and context.