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. 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. 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. Under a closed-set protocol, SENTRY reaches 98.75% accuracy at a 1.25% equal-error rate (EER) on WaveFake, with perfect separation of HiFi-GAN and WaveGlow; its supervised track alone achieves an AUC of 0.9994 (fused: 0.9993). 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%. In our prototype, the feature front end runs as bit-exact integer code on one Arm Cortex-A53 core, and the classifier and fusion stage are synthesized for a Xilinx Zynq UltraScale+ MPSoC using no DSP slices, returning a decision 220 ns after a feature vector is ready. These results indicate that deepfake screening can be placed inline without GPUs or floating-point hardware; generalization to unseen generators is left to future work.
Figures & tables
Fig. 1 : Overview of the proposed framework.
Dataset
Track
Acc. (%)
F1 (%)
AUC
EER (%)
WaveFake
Fused
98.75
98.75
0.9993
1.25
Track A
98.69
98.69
0.9994
1.31
Track B
70.11
70.16
0.7195
29.89
ASVspoof 2019 LA (A01–A06)
Fused
94.75
97.00
0.9884
5.05
Track A
94.17
96.65
0.9948
5.38
Track B
56.83
70.23
0.6024
43.18
TABLE I : Closed-Set Detection Performance on Both Datasets ( 8,000 utterances each, integer datapath). Best value per column in bold.
AI Model
Type
Accuracy of Being AI (%) a
AUC
EER (%)
WaveFake ( 22.05 kHz wideband)
HiFi-GAN
GAN
100.00
1.0000
0.00
WaveGlow
Flow
100.00
1.0000
0.00
Parallel WaveGAN
GAN
99.87
1.0000 b
0.13
MelGAN
GAN
99.36
1.0000 b
0.64
Multi-band MelGAN
GAN
98.43
0.9981
1.57
TABLE II : Detection Accuracy by AI Model (fused score, each generator versus bona fide)
Resource
Used
Available
Utilization (%)
LUT
11,656
230,400
5.06
LUTRAM
1,090
101,760
1.07
BRAM
9
312
2.88
Register
13,249
460,800
2.88
DSP
0
1,728
0.00
TABLE III : Post-Synthesis Resource Utilization of the Classifier and Fusion Kernel (Zynq UltraScale+ XCZU7EV)
Clock (MHz)
Latency (ns)
Throughput (M vectors/s)
Power (W)
100
220
4.55
2.94
TABLE IV : Timing and Power of the Classifier and Fusion Kernel (power is the Vivado on-chip estimate for the whole device, including the processing system)
Method
Front End
Arithmetic
Public multi- corpus eval
Deploy target
CNN–RNN + XAI [ 16 ]
MFCC/Spec
Float NN
No
Cloud
Transformer–CNN [ 15 ]
Mel-spec
Float NN
No
Mobile ∗
RawNet2 [ 18 ]
Raw/sinc
Float NN
Yes
GPU
AASIST [ 8 ]
Raw/graph
Float NN
Yes
GPU
wav2vec2-AASIST [ 19 ]
SSL
Float NN
Yes
GPU
SENTRY (ours)
Integer lifting DWT
Integer (no DSP)
Yes
FPGA Online
TABLE V : Comparison of Design Points with Representative Prior Systems
Audio deepfakes generated by neural text-to-speech and voice-cloning systems threaten speaker verification and public discourse at scale. The core challenge is cross-dataset generalization: detectors trained on one synthesis pipeline collapse on unseen forgeries. We argue that this failure is primarily because of structural synthetic speech artifacts which are multi-timescale trajectory anomalies. Though every existing detector aggregates a fixed-window frame statistics, this misaligns the architecture with the signal. We propose FlowFake, a Liquid Time-Constant (LTC) architecture whose hidden state evolves via a learned ODE, with per-neuron adaptive time constants simultaneously resolving spectral (10ms) and prosodic (2s) cues. At only 34K parameters FlowFake achieves formal BIBO stability and O(dt^4) integration error. On a four-dataset cross domain benchmark (ASVspoof2019-LA, FakeOrReal, InTheWild, MLAAD), FlowFake reaches 75.29% on ASVspoof2019 trained only on FakeOrReal and 79.97% trained only on MLAAD. It outperforms RawGAT-ST and Whisper-DF on every evaluated pair and matching SSL Wav2vec2 (300x larger) at 0.01% of its parameter count. The source code is available on : https://github.com/GhostRider2023/FlowFake
Video-mediated conversational systems rely on synchronized speech and facial video, making audio--visual input integrity important for reliable perception. We study audio-frontend design in a fixed MRDF (Modality-Regularization-based DeepFake)-based pipeline with AV-HuBERT visual processing and an MSCT (multi-scale cross-modal transformer) fusion encoder. We compare spectrogram-style log-filterbank features with frozen wav2vec2.0 and WavLM representations, dual-SSL fusion, executed depth, and shallow-layer aggregation. Across three training runs, full wav2vec2.0 improves mean macro AUROC from 84.19% to 87.19% on generator-disjoint MMDF. Six-layer execution further improves detection, and learned non-uniform aggregation adds 2.53 percentage points, reaching 90.61%. Shallow execution reduces measured inference latency relative to full-depth wav2vec~2.0 across the tested input durations; first-second evaluation reaches 86.85% macro AUROC. In a single-run comparison, dual-SSL concatenation outperforms either single stream, while one-way cross-attention yields lower detection performance and higher latency than concatenation. Layer-weight analysis shows greater emphasis on early speech layers, while removing the selector's visual input has little effect on detection performance.
Audio deepfakes are a growing challenge for the general public, as well as for journalists and fact-checkers. The latter need reliable tools to verify the authenticity of their sources, while at the same time keeping their information private. Commercial deepfake detection solutions rely on cloud-based processing, which raises privacy concerns. To solve this problem, we propose an on-device audio deepfake detection model. We show that a truncated self-supervised backbone with a simple logistic classifier is both very fast and often more accurate than existing solutions. Our solution outperforms the baseline AASIST by 10% and improves inference speed by 40%. We integrate this model into a browser plug-in, which allows journalists and fact-checkers to detect deepfakes easily and securely. Code for the plugin is available at https://github.com/OctavianPascu97/Audio-Deepfakes-Browser-Plugin.
Octavian Pascu, Dan Oneata, Horia Cucu +1
National University of Science and Technology POLITEHNICA Bucharest, Romania · Fraunhofer AISEC, Germany