Yukun Yan, Peng Tang, Rui Chen, Yanru Sun, Cheng Yu, Qilong Han, Jie Zhang, Jing Li · ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 2026
DOI: 10.1145/3799682.3840926
Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).
Differentially private dataset distillation (DP-DD) enables privacy-preserving data sharing by releasing compact synthetic datasets with formal privacy guarantees. However, existing DP-DD methods typically optimize for a fixed images-per-class (IPC) setting, making them ill-suited to heterogeneous users who require distilled datasets with different IPC sizes. Straightforward extensions to support multiple IPC requirements either incur cumulative privacy costs across separate DP-DD runs or sacrifice adaptability by relying on a predefined maximum IPC size. To address this issue, we introduce Size-Adaptive DP-DD, a strategy that decouples private data access from the construction of IPC-specific distilled datasets under a single privacy budget. Building on this strategy, we propose DuetDP-DD , which constructs a reusable DP-synthetic candidate space and derives IPC-specific distilled datasets through tailored post-processing. DuetDP-DD employs a dual-classifier guidance scheme: the Semantic-Anchor Classifier identifies semantically central candidates within each class, while the DP-Classifier provides discriminative guidance, with its feature space supporting diversity-aware selection, thereby preserving class representativeness, inter-class discriminability, and intra-class coverage of distilled datasets across IPC settings. This design allows DuetDP-DD to serve emerging IPC requests with utility-preserving distilled datasets without additional private optimization. Extensive comparative experiments demonstrate that DuetDP-DD achieves competitive utility across IPC settings without incurring the repeated privacy cost of separate fixed-size DP-DD runs. Additional analyses confirm the effectiveness, flexibility, and cross-architecture generalization of DuetDP-DD, underscoring its applicability to privacy-preserving distilled dataset sharing for heterogeneous users.
No comments yet — start the discussion below.