Latent Geometric Chords for Query-Efficient Decision-Based Adversarial Attacks

Researchers introduced Latent Geometric Chords (LGC), a decision-based black-box adversarial attack for computer vision.

The method uses curvature-aware search in latent space plus a residual-based generation scheme to improve query efficiency and preserve image quality.

A variant called LGC-H shows strong transferability, high attack success, and better visual fidelity than prior attacks.

At 5,000 queries, the paper reports SSIM above 0.99 and LPIPS below 0.01, highlighting a subtler threat to robust image models.