InkDrop: Invisible Backdoor Attacks Against Dataset Condensation

Dimensions overall score 7.0

• Existing backdoor attacks against dataset condensation prioritize attack effectiveness and model utility but overlook stealthiness.limitation 90%
• InkDrop enhances the imperceptibility of malicious manipulation in dataset condensation backdoor attacks without degrading attack effectiveness and model utility.method 90%
• InkDrop leverages the inherent uncertainty near model decision boundaries, where minor input perturbations can induce semantic shifts, to construct a stealthy and effective backdoor attack.technical 85%
• InkDrop first selects candidate samples near the target decision boundary that exhibit latent semantic affinity to the target class.method 80%
• InkDrop learns instance-dependent perturbations constrained by perceptual and spatial consistency to embed targeted malicious behavior into the condensed dataset.method 80%
• Extensive experiments across diverse datasets validate the overall effectiveness of InkDrop, demonstrating its ability to integrate adversarial intent while preserving model utility and minimizing detectability.result 85%
• The training objective for the attack model integrates a contrastive InfoNCE loss to align poisoned embeddings with the target class while repelling clean samples.technical 85%
• Dataset condensation introduces a critical vulnerability to backdoor attacks where malicious patterns are implanted into the condensation dataset.technical 90%

Competitive landscape

Competitor map is still being generated for this paper. Enable generation or check back soon.

Startup potential card