AttnDiff extracts LLM fingerprints from differential attention responses to conflicting prompt pairs, summarized via spectral descriptors and compared with CKA, yielding >0.98 similarity for derivatives versus <0.22 for unrelated families.
stress test
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CR 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
AttnDiff: Attention-based Differential Fingerprinting for Large Language Models
AttnDiff extracts LLM fingerprints from differential attention responses to conflicting prompt pairs, summarized via spectral descriptors and compared with CKA, yielding >0.98 similarity for derivatives versus <0.22 for unrelated families.