Deep Fake Categories
Real world modified data: Image, video, sound data is recorded in the real world and then the data is enhanced or altered via SW tools. This includes intentional and malicious insertion or deletion of artifacts in the data stream.
Artificially generated data: GenAI tools (ChatGPT, NanoBanana etc) can be used to generate images, videos and sound data at the click of a button. This technology has a significant positive potential – artificially generating data which is difficult or probably impossible to acquire in the real world (e.g: data for rare genetic diseases which impact ~ 0.001% of the population) and then this data could be used to improve AI SW systems. However, there is also a downside to this technology: Intentional generation of fake and malicious data.
Problem with Deepfakes:
Social media is flooded with malicious and manipulated synthetic data and this data is extensively used to influence public opinion and create chaos in society (Russian Hybrid Attack on Germany.). This trend will only grow in the future and with the widespread and easy availability of Gen AI models.
Scammers are now using manipulated videos to lure people into investment scams. (Financial Scams).
Mitigation:
An important mitigating step in this direction is the EU AI act (article 50, paragraph 4 and other sections) which clearly mandates the requirement to label all visual & audio data generated by artificial means.
C2PA(Coalition for Content Provenance and Authenticity- https://c2pa.org). Most large AI companies insert visible & invisible watermarks to clearly distinguish between synthetic and non-synthetic visual data. However the availability of open source models which can be downloaded and run locally creates a situation wherein malicious actors can and will tweak the model to remove the C2PA watermarking in the generated data.
An alternative approach is the use of tools like SightEngine (https://sightengine.com/detect-ai-generated-images) which are capable of detecting images generated using Gen AI models.
How can EEU contribute:
The EU AI act as published in 2024 is in a very nascent stage. The act specifies certain objectives but the pathway to achieving those objectives is not defined. EEU as a group of experts with significant experience in different engineering domains will contribute towards the concrete definition of the objectives and also the tools & methods used to verify and implement various rules specified in the EU DSA and AI act. As of today, the EU AI act visualizes a scenario where every member country will have processes and organizations to oversee the implementation of the EU AI act. EEU will push for a unified rules and regulations scenario so that companies operating in the EU are not unnecessarily burdened with compliance issues in context of the same laws in different countries.
Of late, many companies have voiced criticism or opposition to the EU AI act. While some of the criticism might be valid, these tactics have been and will be used as a mechanism to weaken these kinds of laws. E.g: Criticism by American companies and individuals against EU Digital Services Act (DSA). EEU will act as a technical pressure group to prevent dilution of rules and regulations in a manner such that it benefits some specific companies or organizations. As a responsible organization we are open to engaging in discussions to address valid concerns in context of the above mentioned laws. However, we will proactively engage and contribute technical facts and data to prevent dilution of rules where there is a possibility of potential harm to individuals and the EU society as a whole.