Zetao Lin, Qiao Yan · ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 2026
DOI: 10.1145/3799682.3840739
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).
Model inversion attacks (MIAs) aim to reconstruct high-quality copies of inaccessible private training data from deep neural networks, thereby posing a severe threat to data privacy. Training-based MIAs are only applicable to reconstructing low-resolution grayscale images. Optimization-based MIAs suffer from complex training of attack models, reconstruction quality degradation caused by iterative optimization, and high query and time overhead. Although diffusion-based MIAs achieve significant advantages in reconstruction quality and generation efficiency, this paradigm still faces two major limitations: 1) Unfiltered public datasets, which severely constrains reconstruction quality. 2) Low execution efficiency, particularly the substantial computational and time overhead incurred by training conditional diffusion models from scratch in the image space. Thus, we propose Lightweight Latent Diffusion Fine-tuning for Model Inversion Attack (LLDF-MI), a novel method that fine-tunes the pre-trained latent diffusion models (LDMs) via Low-Rank Adaptation (LoRA) to achieve efficient and high-quality black-box MIA. Specifically, we first constructs a Top-K positive face fine-tuning dataset via head pose detection and target model confidence filtering for precise visual-identity mappings. On this basis, we design a Multi-Identity LoRA module with a ''partially identity-specific + partially shared'' parameter allocation strategy, combined with conditional encoding and substitute-based identity loss, to precisely guide the generated images toward the target identity feature space. Experiments show that its attack performance significantly outperforms existing black-box and label-only MIAs, and is comparable to advanced white-box MIAs. Furthermore, it substantially reduces the training time overhead of the attack model and effectively improves the generation efficiency of attack images.
No comments yet — start the discussion below.