AutoMIA

LLM agents design new privacy attacks — in natural language.

100 attacks designed autonomously.

Final AUC: 0.73 (+0.19 vs. human baseline)

Automated Membership Inference Attacks: Discovering MIA Signal Computations using LLM Agents
Paper: arXiv Code: GitHub
Explorer
(proposes new ideas)
Idle Iter: 0
    0.500
    TPR@1%: 0.000 TPR@5%: 0.000

    Exploiter
    (refines top performers)
    Programmer
    →
    Executor
    →
    Analyzer
    #id idea AUC
    ... 0 more entries
    0 / 100
    experiments completed
    Best AUC: 0.540
    (baseline: 0.54)
    Human baseline