Differentially Private Motif-Preserving Multi-modal Hashing
The paper presents a novel approach to cross-modal hashing that preserves privacy while enabling efficient retrieval of images and text. The proposed method, DMP-MH, addresses vulnerabilities in existing privacy-preserving techniques by employing a Sanitize-then-Distill framework. Experimental results show that DMP-MH significantly outperforms private baselines while maintaining high performance levels.
- ▪DMP-MH is a framework designed to enhance privacy in cross-modal hashing.
- ▪The method addresses issues related to sensitive behavioral patterns in semantic similarity graphs.
- ▪Experimental evaluations indicate that DMP-MH outperforms existing private methods by up to 11.4 mAP points.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Information Retrieval arXiv:2605.15460 (cs) [Submitted on 14 May 2026] Title:Differentially Private Motif-Preserving Multi-modal Hashing Authors:Zehua Cheng, Wei Dai, Jiahao Sun View a PDF of the paper titled Differentially Private Motif-Preserving Multi-modal Hashing, by Zehua Cheng and 1 other authors View PDF HTML (experimental) Abstract:Cross-modal hashing enables efficient retrieval by encoding images and text into compact binary codes. State-of-the-art methods rely on semantic similarity graphs derived from user interactions for supervision, yet these graphs encode sensitive behavioral patterns vulnerable to link reconstruction attacks.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
