"New Deepfake Threats Loom, Says Microsoft's Chief Science Officer"

Eric Horvitz, Microsoft's chief science officer, testified before the US Senate Armed Services Committee Subcommittee on Cybersecurity, stating that organizations will face new challenges as cyberattacks become more sophisticated, including through using Artificial Intelligence (AI)-powered synthetic media and deepfakes. Deepfakes are described as high-fidelity, synthetic, fictional depictions of people and events made using AI and Machine Learning (ML). They have become a common misinformation tool in the past few years, but new deepfake threats are imminent, according to Horvitz. According to a new research paper from Horvitz, interactive and compositional deepfakes are two emerging classes of threats. Davis Blaloch, a MosaicML research scientist, described interactive deepfakes as the illusion of conversing with a real person. Compositional deepfakes take it a step further, with a malicious actor creating multiple deepfakes to create a "synthetic history." For example, consider fabricating a fictitious terrorist attack, a fictitious scandal, or assembling "proof" of a self-serving conspiracy theory. Blaloch says such a synthetic history could be supplemented with real-world action, such as setting fire to a building. Horvitz stated in the paper that the increasing capabilities of discriminative and generative AI methods are approaching an inflection point. The advancements are providing tools for state and non-state actors to create and distribute persuasive disinformation. Deepfakes will be more difficult to distinguish from reality. Horvitz explained that the problem stems from the Generative Adversarial Networks (GAN) methodology, which is an iterative technique in which the ML and inference used to generate synthetic content is pitted against systems that attempt to distinguish generated fiction from fact. The generator gradually learns to fool the detector. According to Horvitz, with this process at the foundation of deepfakes, neither pattern recognition techniques nor humans will be able to recognize deepfakes reliably. This article continues to discuss the concept of deepfakes, the growing sophistication of deepfakes, and defending against deepfakes.

VB reports "New Deepfake Threats Loom, Says Microsoft's Chief Science Officer"