Differentially Private Motif-Preserving Multi-modal Hashing
A recent study published on arXiv (2605.15460) presents DMP-MH, a novel framework aimed at enabling privacy-preserving cross-modal hashing. This technique encodes both images and text into compact binary codes for streamlined retrieval. However, traditional methods depend on semantic similarity graphs derived from user interactions, which are susceptible to link reconstruction attacks. Existing privacy solutions are inadequate: Differentially Private SGD compromises relational patterns, while graph synthesis techniques encounter Hubness Explosion, where certain nodes disproportionately influence triangle counts, necessitating excessive noise. DMP-MH addresses these issues through a Sanitize-then-Distill approach, which separates privacy concerns from representation learning by limiting sensitivity via deterministic node degree clipping.
Key facts
- Paper arXiv:2605.15460 proposes DMP-MH for privacy-preserving cross-modal hashing.
- Cross-modal hashing encodes images and text into compact binary codes.
- Existing methods use semantic similarity graphs from user interactions.
- These graphs are vulnerable to link reconstruction attacks.
- Differentially Private SGD destroys relational motifs.
- Graph synthesis methods suffer from Hubness Explosion.
- Hub nodes cause single-edge modifications to alter triangle counts by O(N).
- DMP-MH uses a Sanitize-then-Distill framework with deterministic node degree clipping.
Entities
Institutions
- arXiv