Problem
Research questions and friction points this paper is trying to address.
Exposure-Control Attacks
Attribute Inference
On-Device Auditing
Privacy
AccretionLink: On-Device Auditing of Exposure-Control Attacks on Attribute Inference
🤖 AI Summary
This study addresses the challenge of auditing private attribute inference under exposure control attacks by proposing a confidentiality-integrity security game model and an on-device trusted auditing framework. By integrating partial identification modeling for selection probabilities with dependency-aware e-processes, and leveraging Tensor G5 graph computing alongside P-256 signature authentication, the framework achieves efficient verification. Experimental results validate the effectiveness of attack advantage analysis, while successful deployment on the Pixel 10 platform demonstrates fallback-free encoding and key authentication mechanisms. Collectively, this work provides both formal guarantees and practical engineering solutions for ensuring privacy security in on-device AI systems against sophisticated inference attacks.
📝 Abstract
Exposure control lets an adversary rank authentic public posts to strengthen private-attribute inference without altering content. AccretionLink defines confidentiality and integrity games for this attack, models bounded selection odds through partial identification, and constructs dependence-aware time-uniform e-processes. On 52 held-out synthetic profiles, odds-four selection reduced aggregate negative log likelihood at every horizon. At eight posts the advantage was 0.01595 nats (95% CI [0.00890, 0.02336]), three of four target effects survived Holm adjustment, and label-blind model-guided selection caused 6/109 high-confidence false reversals. On 142 PAN15 test profiles, exploratory selection produced a 0.01227-nat advantage but no reversal. A separate TF-IDF selector retained a 0.01470-nat advantage against the unchanged G5 target, while matched identity shuffling did not reproduce it. Pixel 10 encoded all 1,622 held-out posts once with a fallback-free Tensor G5 graph. A P-256 checkpoint authenticated the selected-replay, actual-model, native-report, and operation digests; local KeyInfo identified the signing key as StrongBox-backed.
Innovation
Methods, ideas, or system contributions that make the work stand out.
On-Device Auditing
Exposure-Control Attacks
Time-Uniform E-Processes
Partial Identification
Hardware-Backed Security
🔎 Similar Papers
No similar papers found.
