Assessment of the potential of artificial intelligence technologies and other modern methods for detecting deepfake materials in fraud and other unlawful acts

Abstract

The article assesses how far artificial intelligence and other modern methods can detect deepfake material used in fraud and other unlawful acts. It describes the main forms of synthetic media: face swapping, facial reenactment, voice cloning, full-face synthesis and forged biometric credentials. It then compares the available detection tools: convolutional neural networks (XceptionNet), vision transformers, frequency-domain analysis, biological signal analysis and cryptographic provenance (C2PA). These tools perform almost flawlessly on benchmarks, but their accuracy falls on real-world material. The causes are lossy compression, poor generalisation to unfamiliar data, and vulnerability to adversarial attacks. Recent benchmark evaluations demonstrate that Vision Transformers exhibit superior cross-dataset generalization (11.33% drop) compared to CNNs (>15% drop), while classical machine learning models like Random Forest achieve 99.64% accuracy with 2 ms inference time. The procedural analysis focuses on the law of the Republic of Kazakhstan, covering the admissibility of electronic evidence and the distinct roles of experts and specialists. The article concludes that a detector's probabilistic output cannot stand as independent evidence. It becomes usable only with expert interpretation, explainable AI (XAI) and a verified chain of custody. The proposals include a qualifying feature in Article 190 of the Criminal Code (or an aggravating circumstance in Article 54), mandatory labelling of synthetic content, validated forensic methodologies, an interagency reference database and anti-deepfake safeguards in banking.