10-18, 13:00–13:20 (America/New_York), Track 2 (206a)
Language: English
Real-time deepfake detection has become a critical cybersecurity challenge, as increasingly sophisticated deepfake technologies undermine trust in digital communication and media. Users currently lack effective tools that are integrated directly into their browsers and are unable to quickly verify the authenticity of images they encounter daily. Existing solutions struggle to detect newer, diffusion-generated deepfakes, such as those produced by DALL-E and Stable Diffusion, due to high computational requirements and limitations in detection accuracy. To address these challenges, we introduce VeriPhoto Version 2, a browser-based Chrome extension that leverages DistilDIRE, a lightweight transformer-based model specifically optimized for real-time deepfake detection. Crucially, VeriPhoto Version 2 integrates Apple’s Metal Performance Shaders (MPS) to harness the GPU capabilities present in Mac M1 and M2 chips, enabling rapid, highly accurate detection directly within the browser. This GPU acceleration significantly reduces latency and computational load, making VeriPhoto practical and effective for everyday users to seamlessly identify and combat deepfake images in real-time. This presentation will discuss the journey of how we designed veriPhoto v2 and the hurdles we overcame during its implementation.
Deepfake technologies pose significant threats to authenticity in digital media, undermining trust in both personal and institutional communications. Current detection solutions struggle with this threat due to computational limitations and difficulty identifying diffusion-generated deepfakes produced by models such as DALL-E and Stable Diffusion.
We are introducing Version 2 of VeriPhoto, a Chrome extension designed to detect deepfake images consumed within browsers. The first iteration was presented at HackFest last year, and we were delighted to see substantial community interest. However, usability in Version 1 was hampered by reliance on CPU-based models, resulting in slow processing and lower accuracy rates.
In Version 2, we integrate DistilDIRE, a transformer-based detection model specifically optimized for real-time scenarios. DistilDIRE significantly improves accuracy, especially against sophisticated diffusion-generated deepfakes, outperforming traditional GAN-based methods.
To overcome processing speed constraints, we introduce the novel integration of Apple’s Metal Performance Shaders (MPS). By harnessing GPU capabilities built into Mac M1 and M2 chips, MPS allows VeriPhoto Version 2 to achieve real-time detection within browser environments. This GPU acceleration substantially reduces latency, making the extension practical for daily browsing without significant computational overhead.
Our demo will showcase the DistilDIRE-enhanced VeriPhoto extension operating in real-time, demonstrating its effectiveness and usability improvements.
Finally, we will outline future development pathways, including plans to evolve VeriPhoto into a standalone, multi-environment application and explore further optimizations to enhance performance and accuracy.
Session Breakdown:
-
Introduction to Deepfake Threats and Statistics (3 min):
- Define deepfakes and outline their potential impacts.
- Present recent statistics and knowledge. -
Overview of the Original VeriPhoto Chrome Extension: Strengths and Limitations (3 min):
- Introduction to the VeriPhoto V1 tool and its initial capabilities.
- Discuss technical and practical challenges limiting its accuracy and performance. -
Detailed Exploration of DistilDIRE Model and Integration Benefits (4 min):
- Explain DistilDIRE’s transformer-based architecture.
- Highlight benefits of DistilDIRE integration, particularly its accuracy with diffusion-generated deepfakes. -
Leveraging Apple's MPS for GPU Acceleration and Efficiency (3 min):
- Explain Metal Performance Shaders (MPS) and how it utilizes Apple’s GPU capabilities.
- Detail improvements in processing speed and reduced computational requirements. -
Demonstration: Real-time Deepfake Detection in Browser (5 min):
- Real-time presentation of the VeriPhoto V2 extension enhanced with DistilDIRE and MPS.
- Highlight the speed and accuracy of real-time detection capabilities. -
Future Prospects, Adaptation for Mobile, and Audience Q&A (2 min):
- Discuss potential developments, including stand-alone application.
- Q&A session.
Ishaan is a Security Consulting Analyst based in Calgary with a background in Computing Science and Business. He brings experience in cybersecurity, software development, and data science, along with certifications in cloud, security, and data. Focused on threat detection, incident response, and PEN Testing Ishaan is dedicated to continuous learning and contributing to effective, resilient teams.
