DeepLayer-ID: A Lightweight Multi-Domain Forensic Framework for Real-Time Deepfake Detection in Resource-Constrained...

Area of Science:

• Computer Vision
• Digital Forensics
• Artificial Intelligence

Background:

• Unmanned Aerial Vehicle (UAV) imaging systems are vital for surveillance and monitoring.
• Deepfake technology poses a significant threat to the integrity of aerial sensor data.
• Real-world degradations like motion blur and noise complicate deepfake detection in UAV imagery.

Purpose of the Study:

• To introduce DeepLayer-ID, a novel forensic framework for detecting deepfakes in degraded UAV sensing environments.
• To develop a degradation-aware system that addresses challenges in aerial data integrity.
• To enhance the robustness of deepfake detection under heterogeneous acquisition conditions.

Main Methods:

• Proposed a multi-domain forensic framework decomposing evidence into spatial, frequency, and residual domains.
• Utilized discrete wavelet transform and high-pass residual filtering to capture inconsistencies and anomalies.
• Employed a lightweight transformer-based fusion mechanism for adaptive integration of cross-domain representations.
• Constructed a balanced dataset of 1096 aerial frames with synthetic manipulations and physics-consistent degradations.

Main Results:

• DeepLayer-ID achieved 97.8% accuracy and 0.991 AUC, outperforming existing methods like ResNet-50 and XceptionNet.
• The model demonstrated real-time feasibility with 5.4 M parameters and 9.8 ms inference latency.
• The framework showed superior robustness under various synthetic manipulations and real-world degradations.

Conclusions:

• Structured multi-domain signal decomposition with attention-guided fusion offers a robust solution for deepfake detection in degraded UAV systems.
• DeepLayer-ID provides a computationally efficient and accurate method for ensuring the integrity of UAV visual data.
• The findings highlight the potential of advanced forensic techniques in securing critical aerial sensing applications.