Light-powered AI detects deepfakes at scale with 98% accuracy

The rapid rise of generative artificial intelligence has created a growing challenge for technology companies, media organisations and regulators: how to identify realistic fake videos before they spread misinformation or facilitate fraud.

Now, researchers at the UCLA have unveiled a novel optical AI processor that could substantially improve the speed and efficiency of deepfake detection. The system uses light to carry out part of the computational workload, enabling it to analyse at least 15 video streams simultaneously while maintaining an average detection accuracy of 97.79 per cent.

The technology is described a research paper. Here, the researchers believe the system could form a high-speed first line of defence against the growing volume of AI-generated video appearing across social media platforms, news feeds, surveillance networks and enterprise systems.

What is apparent from the research is that the challenge facing digital platforms is growing rapidly. Notably, advanced generative AI models can now create highly realistic videos using only text prompts, producing content that often lacks the visual flaws and inconsistencies that earlier deepfake detectors relied upon. As synthetic media quality improves, distinguishing authentic footage from fabricated content is becoming increasingly difficult.

At the same time, the volume of material requiring verification continues to expand. Traditional detection systems often process videos sequentially, requiring extensive computational resources for each analysis. As more content needs to be screened, both processing demands and energy consumption increase accordingly. According to the UCLA researchers, many state-of-the-art deepfake detectors require hundreds of billions of floating-point operations for a single analysis. While such systems can be highly accurate, scaling them to safeguard massive video platforms remains a significant challenge.

Letting light do the computing

The UCLA team, led by Professor Ozcan, approached the problem from a different direction. Rather than relying exclusively on electronic processing, they designed a hybrid digital-optical architecture in which some of the neural network computations are performed physically through the propagation of light. The system first employs a lightweight digital encoder that extracts key spatial, temporal and spectral characteristics from each video. This compressed information is converted into optical phase patterns and projected onto a programmable spatial light modulator.

The resulting light waves then pass through a passive optical decoder, where diffraction and light propagation effectively perform calculations that would otherwise require substantial digital processing power. At the output stage, optical detectors generate authenticity scores for each video stream. In practical terms, the approach replaces part of a computationally intensive digital neural network with an optical process that can evaluate many videos in parallel. The work is part of a broader effort to advance a field that seeks to use the physics of light to perform calculations more efficiently than conventional electronic hardware.

The researchers tested the processor using the widely used https://cse.buffalo.edu/~siweilyu/celeb-deepfakeforensics.html deepfake benchmark dataset. During each optical pass, the system examined 15 videos simultaneously and achieved an average detection accuracy of 97.79 per cent. It also demonstrated a sensitivity of 99.86 per cent, meaning it identified almost all manipulated videos correctly. For a screening tool, this metric may be particularly important. Missed detections can allow manipulated content to pass through moderation systems unchecked, increasing the risk of misinformation or malicious activity.

The exceptionally high sensitivity translated into a false-negative rate of approximately 0.14 per cent, suggesting that only a tiny proportion of deepfake videos were incorrectly classified as genuine. Even when researchers increased throughput to 18 simultaneous video streams, the system maintained an accuracy rate above 96 per cent. Such results indicate that optical parallelisation may offer a practical route for screening vast quantities of video content without proportionally increasing computational overhead.

Greater capability without greater energy costs

The study also demonstrated another potential advantage of optical computing. The researchers found that adding extra passive diffractive layers to the optical decoder improved performance on more challenging deepfake manipulations. Incorporating two additional layers increased detection accuracy by around 6.8 per cent. Unlike larger digital neural networks, these optical elements require no additional electrical power during inference. Once fabricated, they continue to perform computations through the natural behaviour of light. This suggests that future optical AI systems could potentially increase analytical capability without incurring the substantial energy costs typically associated with larger digital models. As concerns grow over the environmental footprint of artificial intelligence, energy-efficient AI architectures are attracting increasing interest from both industry and academia.

Tested against Google’s Veo 3

The team also explored whether the optical processor could handle newer forms of AI-generated video. Rather than focusing solely on traditional face-swapping deepfakes, researchers evaluated the system using content generated by Veo 3. Such videos often pose a greater challenge because they do not necessarily contain the artefacts associated with older deepfake techniques.

Despite this, the optical-neural processor achieved an accuracy of 94.8 per cent and a sensitivity of 97.61 per cent after only limited fine-tuning. The findings suggest the underlying approach may remain effective as generative video technology continues to evolve.

Beyond speed and efficiency, security represents another notable aspect of the research. The UCLA team reported that the system demonstrated resilience against black-box adversarial attacks and offers inherent protection against certain white-box attacks. Because part of the model exists physically within the optical hardware, attackers may find it significantly more difficult to reconstruct or manipulate the detection process. The processor also maintained strong performance when videos were affected by image noise, compression artefacts, blur and alignment variations that commonly occur in real-world environments. According to the researchers, these results point towards a potentially important advantage of optical AI systems: the ability to combine computational efficiency with greater robustness against attempts to deceive or circumvent machine-learning models.

The researchers are not proposing that optical processors replace digital deepfake detectors altogether. Instead, they envision a layered approach. Large volumes of video could first be screened by a fast, highly sensitive optical processor. Content identified as potentially suspicious could then be routed to more sophisticated digital forensic systems for detailed analysis. Such a workflow could combine the strengths of both approaches. Optical processing offers speed, scalability, lower energy requirements and resistance to attack, while conventional AI systems can provide more detailed forensic assessment when required.

As AI-generated content continues to proliferate, hybrid architectures of this type may become increasingly attractive. They offer the prospect of detecting manipulated media at scale without the escalating computational costs that accompany many current approaches. For platforms struggling to maintain trust in an era of synthetic content, the UCLA team’s work suggests that one of the most powerful tools for combating deepfakes may not be a larger neural network at all, but the computational capabilities inherent in light itself.