Multimodal Enterprise Authentication Resistant to AI-Based Identity Manipulation: Facial Anti-Spoofing and Multilingual Deepfake Speech Detection
Year: 2026
Type: conference-paper
Abstract: The accelerating proliferation of generative artificial intelligence has transformed deepfake-based identity manipulation into a credible, scalable, and operationally viable threat against enterprise authentication systems. Modern face-swap pipelines can seamlessly superimpose any target identity onto an attacker’s live video with fidelity sufficient to defeat human inspection, while neural text-to-speech engines enable accurate voice cloning from as few as three seconds of reference audio. This work presents a production-grade multimodal authentication platform that jointly performs facial liveness anti-spoofing and multilingual synthetic speech detection. The facial module employs a ResNet-18 convolutional neural network fine-tuned on FaceForensics++ and Celeb-DF v2 datasets. The audio module uses a lightweight CNN classifier operating on MFCC spectrograms trained on ASVspoof 2019 augmented with in-house Kazakh and Russian TTS corpora totalling 40,080 utterances from 22 distinct synthesis systems. Score-level fusion with learned weight α*=0.55 produces a three-tier decision: Genuine, Suspicious, or AI-Generated. On a 4,800-session held-out test, the system achieves AUC 97.4%, EER 4.1%, Accuracy 94.9%, and F1-score 0.961. The prototype delivers 1.74 s mean end-to-end latency with OAuth 2.0 enterprise IdP integration.
Institution: Astana IT University
Language: English
Subfield: Signal Processing
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7203608917