4 papers
When T2I Synthetic Data Backfires: Amplified Privacy Risks in Real-Synthetic Mix Training
Na Li, Boyu Kuang, Hongsheng Hu +4
The paper shows that mixing text‑to‑image synthetic data with real data during training can increase privacy leakage of the real samples, and introduces a method (RSMixLeak) to mea…
From Multiplicity to Vulnerability: Privacy Amplification Risk from One-Dataset-Multiple-Model Exposure
Qirui Huang, Na Li, Hongsheng Hu +3
To efficiently exploit a valuable data source (e.g., facial or medical images), it is frequently harnessed to fulfill multiple learning objectives (e.g., facial recognition, age es…
CompLeak: Deep Learning Model Compression Exacerbates Privacy Leakage
Na Li, Yansong Gao, Hongsheng Hu +2
Model compression is crucial for minimizing memory storage and accelerating inference in deep learning (DL) models, including recent foundation models like large language models (L…
From Pixels to Trajectory: Universal Adversarial Example Detection via Temporal Imprints
Yansong Gao, Huaibing Peng, Hua Ma +5
For the first time, we unveil discernible temporal (or historical) trajectory imprints resulting from adversarial example (AE) attacks. Standing in contrast to existing studies all…