Deepfakes are digital-media which may contain audio, videos or images synthesised by usage of Generative Adversarial Networks (GANs) with the aid of Artificial Intelligence (AI) technologies. These deepfakes have the capability to replicate and mimic the behaviour of real people. Although deepfakes are beneficial in filmmaking and in education sector, but they can have serious implications within the other areas or fields such as in politics, social network platforms, human security and law. The deepfakes can mislead several people with false information. By creating false evidence that can behave like real people and damage the reputation of one. There are several ways for deepfake detection. They work on the basis of analysing and discriminating between various features such as facial features, movements, blinking, variations in voice, variation in tones and background noises. However, these detection systems may be vulnerable to quantum adversaries and futuristic adversarial attacks. This may result in an ambiguity within trustworthy detection systems, which creates an overriding necessity for effective and trustworthy detection framework. The paper highlights a reliable deepfake image detection framework based on the ResNeXt architecture optimized with use of lattice-based adversarial training that is learning with errors (LWE) mechanism to make it resilient against several adversarial manipulations. In addition to this, when followed by the unification of Kyber and Dilithium with quantum cryptography methods, these ensure the authenticity and encryption of the detection results. The proposed scheme DeepQShield is quantum-resistant because it incorporates the executions of post cryptography algorithms and is trained and tested on the Deepfake Detection Challenge dataset (DFDC). On the DFDC database it achieved significantly higher accuracy of 99.28% and an impressive AUC value of 0.9997. When compared to the existing systems such as EfficientNet-B7 (accuracy: 97.2% on DFDC), Vision Transformers (ViT) (90 to 98% on Celeb-DF and DFDC), Multi-attentional CNN-LSTM networks (98.2% on DFDC), FuzzyDFD (accuracy: 99% FF++ and 93% on (Celeb-DF). DeepQShield outshines the conventional models in terms of security, scalability, accuracy and robustness making it best suitable for various applications in real-world scenarios like face forensics, social media data authentication.
Shreeya et al. (Fri,) studied this question.