Evidence · DiggingBeagle record
GARD purifies poisoned CLIP encoders but small contaminated repair corpus reactivates attack
Original GARD backdoor-repair paper tests image/text encoders under clean and contaminated reference sets. In one BadCLIP result, two injected poisoned repair examples increased measured attack success from 0.82% to 68.68% without a comparable clean-accuracy warning.
- Source role
- primary disclosure
Evidence record
Original GARD backdoor-repair paper tests image/text encoders under clean and contaminated reference sets. In one BadCLIP result, two injected poisoned repair examples increased measured attack success from 0.82% to 68.68% without a comparable clean-accuracy warning.
Claim-level citations (1)
- supportsA few poisoned repair images could restore a CLIP backdoor after GARD purification: In the paper’s BadCLIP experiment, adding two poisoned examples to the GARD repair data increased measured backdoor attack success from 0.82% to 68.68%.
Section 5, reference-set size and contamination/backdoor reactivation experiments