DEFORM

Acronym: DEFORM
Title: Deepfake detection and Explainability using FOundation Representation Models

What problem does DEFORM aim to solve, and why are current forensic tools struggling to keep up with deepfakes?
Deepfakes refer to AI-generated or manipulated content that mimics real persons, objects, places or events and falsely appears authentic. According to Europol, deepfakes have gone from a niche curiosity to a systemic risk, and are now used in organised, cross-border crime.

Forensic examiners have to answer one question: is this recording real? Their tools were not built for it. Most are trained to spot a single kind of fake, so new techniques slip past. They handle images, audio and video separately. And they lose accuracy on the material that actually turns up in cases — footage that has been compressed, re-uploaded or poorly recorded. Crucially, they cannot explain why something was flagged, which leaves examiners unable to justify conclusions in court. 

DEFORM is building what is missing: one system covering images, audio and video, able to adapt as new techniques appear, tested by police and forensic institutes on realistic material and designed to explain its findings in a way that lawyers, judges and juries can follow and challenge.
 

How will DEFORM detect and trace deepfakes and AI-generated or manipulated audio, video, and images?
DEFORM is developing a modular framework that handles audio, images and video in one system rather than three disconnected tools.

Detection is built on foundation models, large, general-purpose AI models, combined with few-shot learning, which allows the system to recognise a new manipulation technique from a small number of examples rather than requiring full retraining. This is what enables the framework to keep pace with a threat that evolves continuously, and it is reinforced by a new benchmark dataset of deepfakes generated under documented, reproducible conditions, with balanced demographic representation to support fairness testing.

Beyond detection, DEFORM addresses the question of source. Different generation technologies leave different traces, and tracing content back to the family of technology that produced it strengthens its evidential value, while also allowing investigators to establish links between otherwise unconnected cases.
Multi-level explainability runs through the whole framework: rather than flagging content, the system localises what was manipulated and generates explanations tailored to the audience, from forensic peers to legal professionals to lay audiences
 

What specific role and expertise does VUB contribute to the DEFORM consortium??
VUB leads a  workpackage on the Legal, Ethical and Societal Framework, which keeps the technology aligned with EU law and fundamental rights. Specifically, VUB identify the legal, ethical and societal values affected by deepfakes under the GDPR, AI Act, LED and the Charter. Moreover, we will analyse how AI-based detection interacts with fundamental rights in criminal proceedings (fair trial, equality of arms, presumption of innocence, non-discrimination) including the admissibility and probative value of detection outputs as evidence across five jurisdictions. VUB also delivers the Fundamental Rights Impact Assessment framework . The reserach is carried out by the Fundamental Rights in a Changing World Lab (FRCL).

How will the project ensure that its tools can be used reliably in real-world police investigations and legal proceedings across Europe?
DEFORM approaches this from four directions at once.

Practitioners shape the work rather than receive it: Four police authorities and forensic institutes  (NFI, FRMOI, PIT and POLP) sit inside the consortium as full partners, defining operational requirements from the outset and carrying out the testing themselves.

Validation takes place in operational environments: The FRMOI WorkBench and the NFI Hansken platform reproduce the complete forensic process, from analysis through to case reporting, using realistic degraded material of the kind examiners actually encounter. Maturity progresses in two measured stages: technology validated in a relevant environment  and demonstrated in realistic policing contexts.

Legal alignment is designed in, not retrofitted: The framework is mapped throughout against the AI Act, the GDPR, the Law Enforcement Directive, the Charter of Fundamental Rights and the international forensic standards ISO/IEC 21043:2025, 17025 and 27037. A comparison of criminal procedural law across Belgium, France, Italy, the Netherlands and Norway ensures that what the system produces can withstand scrutiny in different national courts, not only one.

Scrutiny and uptake come from beyond the consortium: An Independent Advisory Board bringing together judiciary, legal experts and civil society reviews progress throughout the project, and Europol’s Innovation Lab contributes to both requirements and validation. By the project’s end, more than fifty practitioners will have been trained, and the WorkBench will be released open-source through Europol’s Tools Repository,  placing the results directly in the hands of forensic laboratories across Europe.