The Generative AI (GAI) Academic Advisory Group (AAG) is an informal, voluntary advisory group where members provide their expertise on deepfake technology and the threat it poses to UK law enforcement. The AAG was established to support a wider cross-law enforcement response to deepfakes. Members of the AAG are academics with a range of research interests related to GAI and policy leads for several GAI threats areas. Over 12 months, the AAG will meet online each quarter and an insight into the discussions from each meeting will be provided. 

This was the third AAG meeting. The group heard from colleagues in the Home Office on their Deepfake Detection R&D Programme and researchers from the Centre for National Training and Research Excellence in Understanding Behaviour (CENTRE-UB) who presented their PhD research - Bio-behavioural and Perceptual Approaches to Enhance Deepfake Analysis. Lastly, the National Police Chiefs’ Council (NPCC) Child Sexual Abuse Material (CSAM) lead and Dstl colleagues jointly presented findings on workflow insertion points for AI detection capabilities and a testing and evaluation framework for detection tools. 

Following the Deepfake Detection Challenge, the Home Office have finalised their report and will be looking to address their recommendations in the next financial year. Plans are being developed for a showcase event on deepfake detection at Security and Policing later this year to drive collaboration with the international community. This is aligned to wider Home Office intentions to develop a benchmarking capability for deepfake detection tools. A benchmarking capability would support the assessment and identification of promising international deepfake detection tools for greater operational testing in the UK. 

Researchers from the CENTRE-UB provided an overview on their proposed approaches to develop a model capable of detecting deepfakes by learning patterns of inconsistency in facial movements and dynamics. This work will use bio-behavioural cues to support the identification of spatiotemporal patterns, and subtle physiological and emotional changes that are unique to humans, to be reliably used to detect deepfakes in an automated fashion. The group raised some useful points around the importance of the quality of the training data, the need to consider bias and explainability in the model’s outputs and positioning this work against advances in AI. NPCC is a collaborative partner on this PhD and will look to support with real-world use case testing at later stages of this work. 

Following this, the NPCC CSAM lead spoke about findings from a recent project looking at workflows in child sexual abuse investigations and insertion points for AI image detector capabilities. An overview of the sophistication and scale of the problem was provided. Criminal capabilities to generate a high volume of realistic CSAM has improved significantly in the last few years. Within the investigative workflow, the greatest challenge identified was the need for tooling to support the extraction and assessment of data from seized devices. At this stage, images need to be prioritised and graded before identifying any potential victims. As such, overall points of intervention were identified, and further work will follow to determine the intervention mechanisms. 

Finally, a colleague from Dstl talked through results from the evaluation of image and audio detection models. Firstly, a testing and evaluation procedure was created to understand the effectiveness of detection tools in specific policing use cases, which were detecting AI-generated CSAM images and more general deepfake audio. This procedure was then used to evaluate the synthetic image and audio detection models created by Dstl. The robust testing and evaluation framework considered overall performance, fairness and bias, and computational efficiency. A discussion followed on the results of the models, in particular their ability to not be biased by a range of demographic and protected characteristics. Going forward, there is an ambition to build upon this testing procedure as new deepfake tooling emerges.