Argus AI framework detects privacy leakage across social media posts
New agentic framework uses abductive reasoning to piece together scattered clues from multiple posts.
Public social media posts can leak private information when seemingly harmless cues across text, images, and metadata are combined. To address this, researchers from multiple institutions propose SopriBench, a synthetic benchmark derived from leaked patterns in Rednote and Instagram accounts. SopriBench covers 50 user profiles and 1,569 images with attributes like contextual sensitivity, granularity, and inference difficulty. They also introduce the Privacy Exposure Score (PES), a metric that weights value granularity by contextual sensitivity to capture exposure severity beyond binary accuracy.
To automate leakage inference, the team introduces Argus, a training-free agentic framework inspired by abductive reasoning. Argus forms hypotheses from accumulated evidence, verifies supporting cues, and aggregates cross-post information into privacy profiles. It achieves a PES of 0.55, a 25% improvement over the strongest baseline, with the largest gains on cross-post leakage. This work highlights how AI agents can reconstruct sensitive user profiles from multimodal social media data, raising important privacy and security implications.
- SopriBench includes 50 user profiles and 1,569 images with attributes for multimodal privacy leakage research.
- Argus uses abductive reasoning to infer private attributes from cross-post evidence without additional training.
- Argus achieves 0.55 PES, a 25% improvement over baselines, with strongest gains on cross-post leakage.
Why It Matters
This framework shows how AI can reconstruct sensitive profiles from scattered social media posts, raising urgent privacy redesign needs.