You Are Who You Know and How You Behave: Attribute Inference Attacks...
summary
Presents attribute-inference attacks that combine two data sources — a user’s social network friends and their behavioral records (e.g., liked pages, app reviews) — to infer private attributes like home city, occupation, and interests. Introduces a novel integrated model and attack algorithms, with theoretical analysis and experiments on a 1.1M-user dataset showing much higher accuracy than prior single-source attacks (e.g., 57% city inference overall, >90% on a selected half via confidence estimation).
AI-written summary · about
chapters (22)
- 0:00Intro
- 0:34Online Social Networks are Popular
- 1:01OSNs Are a Mixture of Public and Private Information
- 2:23Existing Attribute Inference Attacks
- 3:45Roadmap
- 3:58Threat Model
- 5:27Our Attack Algorithm: High-Level Overview
- 6:26Social-Behavior-Attribute (SBA) Network
- 8:02Vote Distribution Attack (VIAL) Algorithm
- 8:59Phase - Distributing Vote Capacity
- 9:29Local Rule I: Dividing
- 10:27Local Rule II: Backtracking
- 11:09Local Rule III: Aggregating
- 11:40Phase II
- 12:21Evaluation Data
- 14:29Evaluation Setting
- 15:00Comparing with (Best) Friend-based and Behavior-based Attacks
- 15:59Backtracking is Important
- 16:47Sharing More Behaviors Makes You More Vulnerable
- 17:30Confidence Estimation
- 18:23Trade-off Result
- 19:03Conclusion
related talks
26mAttriGuard: A Practical Defense Against Attribute Inference Attacks via Adversarial Machine Learning
39mA Run a Day Won't Keep the Hacker Away: Inference Attacks in Fitness Tracking Social Networks
01h 04mPrivacy Dependencies
40mEvasion, Poisoning, Extraction, and Inference: Tools to Defend and Evaluate
41mHow Attackers Reconstruct You: Measuring Identity Exposure Across Four Attack Surfaces
51mDiscover how your generic data makes you unique among millions of users
35mDo AI-models gossip?
01h 43mPerspective