DMP-MH preserves privacy in multi-modal hashing with 92.5% accuracy

Zehua Cheng, Wei Dai, and Jiahao Sun introduced DMP-MH at arXiv to address privacy vulnerabilities in cross-modal hashing. Current methods rely on user-interaction graphs to guide binary code learning, but those graphs expose sensitive behavioral patterns to link reconstruction attacks. Existing privacy techniques fail: differentially private SGD destroys structural motifs by treating samples independently, and graph synthesis suffers from 'Hubness Explosion'—where hub nodes in scale-free networks cause single-edge edits to alter triangle counts by O(N), requiring enormous noise. DMP-MH decouples privacy from learning via a Sanitize-then-Distill framework. First, it clips node degrees deterministically to bound L2-sensitivity of triangle motifs independently of dataset size. Then, it generates a sanitized synthetic graph using Noisy Mirror Descent under (ε,δ)-Edge DP. Finally, dual-stream hashing networks distill this topology with a holistic structural loss that enforces cross-modal alignment.

Evaluated on MIRFlickr-25K and NUS-WIDE under strict inductive settings, DMP-MH achieved up to 11.4 mAP points improvement over private baselines and retained 92.5% of non-private performance. The approach effectively eliminates the need for prohibitive noise injection, making it suitable for large-scale private retrieval systems. The paper includes full algorithmic details and ablation studies, and the code is expected to be released. This work is particularly relevant for applications in medical imaging, social networks, and any domain where user interaction patterns must remain confidential while enabling fast multi-modal search.