Skip to content
InfoResearchPreprint

DetectViT: Test-time Backdoor Detection for Vision Transformers via Inter-Head Attention Discrepancy

Published
Record updated
View JSON

Summary

DetectViT is a test-time backdoor detection method for Vision Transformers that needs no training data, model outputs, or trigger knowledge. It identifies two attention anomalies caused by backdoor triggers, complete and partial attention hijacking, and scores them with an inter-head consistency score and an inter-head entropy variance score. Thresholds come from as few as 64 to 128 out-of-distribution calibration samples, and on BadCLIP it reports a 100% true positive rate and a 0.17% false positive rate.