SENTRY: A Streaming Edge Network Trust Engine for Real-Time Audio Deepfake Detection

Key points

• This paper presents SENTRY, a hardware-oriented, streaming trust engine for inline deepfake detection at the network edge.1
• SENTRY employs an integer lifting-wavelet front end together with a bank of multiplier-free transforms, including wavelet-packet envelopes, Walsh-Hadamard sequency profiles, waveform micro-statistics, and long-term spectral features, designed so that acoustic analysis needs only integer add, shift, compare and bitwise operations.1
• A supervised gradient-boosted tree classifier is combined with an unsupervised Mahalanobis out-of-distribution detector trained solely on authentic speech, and their outputs are fused by a lightweight logistic layer to produce graduated LOG, CHALLENGE, or DROP actions.1
• On ASVspoof 2019 LA (attacks A01-A06) over an 8 kHz mu-law telephony channel, it reaches 94.75% accuracy at a 5.05% EER, with a supervised-track AUC of 0.9948 (fused: 0.9884) and per-attack EERs below 7%.1

Sources (1)

• [1]SENTRY: A Streaming Edge Network Trust Engine for Real-Time Audio Deepfake Detection
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 17, 05:22 AM

  Real-time voice cloning has enabled large-scale streaming deepfake and vishing attacks, while existing Network Detection and Response (NDR) solutions rely on GPU-based post-processing that requires buffered audio and generalizes poorly to unseen speech synthesis models.

  This paper presents SENTRY, a hardware-oriented, streaming trust engine for inline deepfake detection at the network edge.

Extractive summary: sentences quoted from the sources.