Browse Teach Curate

Steganography-Based Sample-Specific Backdoor Attacks on On-Device Deep Learning Models

A backdoor attack against the real deep-learning models embedded in mobile apps that hides its trigger with DNN-based image steganography, producing imperceptible, per-sample triggers instead of a fixed visible patch. By extracting on-device models from real apps and injecting these stealthy triggers, it achieves higher attack success than prior patch-based methods while preserving clean-input accuracy.

K-2SO Text

JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 1 Stealthy Backdoor Attack to Real-world

A backdoor attack plants a hidden trigger in a deep-learning model so that inputs carrying the trigger are misclassified to an attacker-chosen label while normal accuracy is preserved. This work targ…