Skip to main content

AI Detection and Content Authenticity

This 180-day program is divided into 15-day blocks focusing on deepfake/media forgery detection, watermarking techniques, and content provenance. Each block lists the topics covered and the practical deliverables. A comprehensive tech stack for this module is provided at the end.

Block 1 (Days 1–15): Foundations of AI-Generated Content & Deepfakes​

Topics Covered: Overview of AI-generated content and threats to authenticity; types of deepfakes (image, video, audio) and their societal impacts; introduction to media forensics and detection challenges (e.g., issues of transferability and robustness in detectors). Familiarization with notable incidents of deepfake misuse and the motivation for content authenticity tools.

  • Deliverables: A research brief summarizing deepfake technology and its implications (including recent real-world examples); an internal presentation outlining key challenges in deepfake detection (e.g. generalization to new fakes, interpretability ); environment setup for subsequent labs (installation of deepfake datasets and basic forensic tools).

Block 2 (Days 16–30): Deepfake Detection Techniques (Images & Video)​

  • Topics Covered: Classical and AI-based techniques for detecting deepfake images and videos. Visible artifacts (e.g., facial inconsistencies, blinking issues in early deepfakes), metadata analysis, and modern deep learning approaches using convolutional neural networks and temporal features. Study of state-of-the-art models from academic literature and benchmarks (e.g., methods from the DeepFake Detection Challenge). Examination of emerging real-time detection methods like Intel’s FakeCatcher, which analyzes subtle blood flow signals in videos.
  • Deliverables: Hands-on analysis of a deepfake video dataset (e.g., FaceForensics++); implementation of a simple deepfake detection model or use of an open-source detector (PyTorch/TensorFlow) on sample videos. A short report evaluating the detector’s accuracy and failure cases. Bonus: a demo showing near real-time detection on a video stream, referencing how FakeCatcher achieves 96% accuracy via physiological cues.

Block 3 (Days 31–45): Deepfake Detection Tools & Project​

  • Topics Covered: Tools and libraries for deepfake detection and creation to better understand both offense and defense. Introduction to popular deepfake creation tools (DeepFaceLab, FaceSwap) to generate controlled fakes for testing. Exploration of detection toolkits and pretrained models (e.g., OpenCV’s dnn modules, academic GitHub projects for deepfake detection). Emphasis on dataset curation and evaluation metrics (precision, recall, ROC curves) for detection systems.
  • Deliverables: A mini-project where trainees create a controlled deepfake (e.g., swapping a face in a video) and then attempt to detect it using a chosen algorithm. Deliver a comparative analysis of at least two detection techniques on these fakes (e.g., one classical forensic method vs one deep learning method), with performance metrics plotted. Code repository of detection experiments and a brief video demonstration of the detection in action.

Block 4 (Days 46–60): Digital Watermarking & Resilience Fundamentals​

  • Topics Covered: Introduction to digital watermarking for images, video, and text. Types of watermarks: visible vs invisible, fragile vs robust. Study of algorithms such as DWT (Discrete Wavelet Transform) and DCT (Discrete Cosine Transform) based watermarks commonly used in image watermarking. Understanding how modern generative models like Stable Diffusion automatically embed invisible watermarks in output images. Discussion of text watermarking proposals for AI-generated text (e.g., watermarking large language model outputs) and their current limitations.
  • Deliverables: Lab exercises embedding invisible watermarks into sample images (using libraries like invisible-watermark used by Stable Diffusion) and then detecting them. Experiments to test watermark resilience: apply transformations (cropping, noise, compression) to watermarked images and see if the watermark survives. Document the results in a report on watermark robustness versus common image manipulations.

Block 5 (Days 61–75): Watermark Attacks & Robustness​

  • Topics Covered: Advanced study of watermark resilience and attacks. Techniques for removing or disrupting watermarks (cropping, resizing, filtering attacks) and evaluating their success. Exploration of robust watermarking algorithms that withstand such attacks better (overview of research on improved watermarking beyond DWT-DCT). Case studies: how malicious actors attempt to defeat watermarks in AI content, and how detection can be made more robust. Introduction to watermarking in video (challenges of temporal consistency) and audio watermarking.
  • Deliverables: A controlled experiment in which trainees try to break the watermarks from Block 4: e.g., write a script to crop or rotate images and attempt to remove the embedded “StableDiffusionV1” marker. Measure the success (using watermark decoder to see if marker still detectable). Deliver a brief report on which attacks succeeded and which the watermark withstood, with suggestions for improving watermark algorithms. Optionally, implement a more robust watermark scheme from literature and test it.

Block 6 (Days 76–90): Copyright Traceback & Content Fingerprinting​

  • Topics Covered: Techniques for tracing the origin of content and detecting copyright infringements. Introduction to content fingerprinting algorithms (perceptual hashes like pHash, Facebook’s PDQ, YouTube Content ID) that identify duplicate or derivative media. Methods for image/video reverse search to find originals (e.g., using TinEye or Google Lens). Legal and technical aspects of copyright metadata and tracking. Understanding how persistent watermarks or fingerprints can enable traceback of AI-generated vs original content. Discussion of real-world systems (YouTube’s Content ID) and their accuracy.
  • Deliverables: Creation of a content fingerprinting pipeline: generate perceptual hashes for a set of images or videos and use them to match against a modified set. Deliver a tool or script that, given a query image, finds potential source matches from a dataset (simulating a copyright traceback). Also, produce a short case study (1-2 pages) of a known incident where AI-generated content raised copyright issues and how fingerprinting or tracebacks could help (citing any relevant real cases or research).

Block 7 (Days 91–105): Content Provenance Standards (C2PA)​

  • Topics Covered: In-depth look at the Coalition for Content Provenance and Authenticity (C2PA) standard for media provenance. C2PA goals and architecture: attaching a tamper-evident manifest to images/videos that records who created or edited the content . Study of C2PA’s components (claims, assertions, cryptographic signatures, manifest store). How C2PA uses cryptographic Public Key Infrastructure (PKI) to certify content history . Review of current adoption (e.g., Adobe’s Content Credentials embedding provenance in Photoshop exports, recent support in OpenAI’s DALL·E 3 ).
  • Deliverables: Read and summarize key parts of the C2PA technical specification (focusing on how a manifest is structured). As a hands-on deliverable, use an open-source C2PA tool or SDK (e.g. the official C2PA Rust or JavaScript SDK ) to attach a simple provenance claim to an image (e.g., “Alice created this image on X date with Y tool”) and then verify it. Provide the image file with the embedded manifest and a verification log showing the signature validation.

Block 8 (Days 106–120): Provenance Tooling & Signed Content Embedding​

  • Topics Covered: Practical tools for content signing and verification. Installation and use of C2PA open-source SDKs and command-line tools for embedding manifests into media (images and video). Exploring signed content embedding beyond C2PA: e.g., camera hardware signing (Nikon’s in-camera signing concept), and Project Origin (an initiative by media companies for authentic capture). Also, consider complementary approaches like using blockchain or NFTs to record content provenance (pros and cons). Real-world examples of signed content pipelines in journalism and law enforcement.
  • Deliverables: Develop a small Provenance Embedder tool: a script that takes an image and a description of edits, and produces a C2PA-signed image. Conversely, develop a Verifier script that reads a given image’s C2PA manifest and outputs the history (who signed, what edits claimed). Test these on a set of edited images (e.g., original -> Photoshop edit -> final, with manifests at each step). Deliver the code and a usage guide. If possible, simulate a tampering scenario (alter the image file without updating manifest) and show that verification fails (tamper detection).

Block 9 (Days 121–135): Real-Time Deepfake/AI Content Detection​

  • Topics Covered: Applying detection and provenance in real-time scenarios, such as live video streams. Challenges of real-time deepfake detection: low latency requirements and high accuracy to avoid false alarms. Study of recent advancements like Reality Defender’s real-time video call deepfake detector (plugin for Zoom). Understanding how streaming frameworks (e.g., FFmpeg, GStreamer) can integrate ML models for on-the-fly analysis. Introduction to research on live video provenance: e.g., a proposed system using C2PA manifests in HTTP Live Streaming to sign each segment with minimal delay.
  • Deliverables: Prototype a real-time detection setup: for example, use a webcam feed or sample video stream and process frames on-the-fly with a lightweight deepfake detection model. The deliverable is a running demo (and code) that flags video frames as authentic or fake in real-time (even if using a simple heuristic due to resource limits). Additionally, write a one-page memo on how one could integrate C2PA manifests into a livestream (referencing the idea of embedding provenance in HLS segments with TPM signing for security).

Block 10 (Days 136–150): Integrated Content Authenticity Pipeline​

  • Topics Covered: Bringing together detection, watermark, and provenance into a cohesive workflow. Designing a pipeline where incoming content (image or video) is first verified for a C2PA manifest and signature (provenance check), then analyzed for known watermarks (AIgeneration markers), and finally run through deepfake detectors if authenticity is still in doubt. Discuss pipeline engineering considerations: performance optimization (e.g., only run heavy deepfake analysis if simpler checks fail), and usability in real-world settings (plugins for social media or messaging to warn users of fake content).
  • Deliverables: Design and implement a simplified Content Authenticity Toolkit – a script or set of scripts that take an input image/video and produce an authenticity report. The report should include: whether a valid C2PA signature was found (and details if yes), whether an invisible watermark was detected (and its message if decoded), and the output of a deepfake detector model. Deliver the code and sample reports on a few test inputs (e.g., one genuine photo with C2PA, one AI-generated image with a known invisible watermark, one deepfake video with no provenance). Also provide system architecture documentation (flowchart of the pipeline).

Block 11 (Days 151–165): Evaluation and Red Teaming​

  • Topics Covered: Testing the limits of the developed systems. Evaluation methodologies for authenticity tools: false positive/negative analysis, robustness to adversarial attacks. “Red teaming” exercises where trainees attempt to fool the detection pipeline (e.g., by lightly modifying deepfakes to evade detectors, or stripping/forging provenance metadata). Incorporating feedback from these tests to harden the system. Discussion of ethical considerations and the arms race between deepfake creators and detectors. Review current research frontiers (GAN-generated deepfakes that try to avoid detection, new forensic techniques like heart-rate detection in videos, etc.).
  • Deliverables: A formal evaluation report of the toolkit from Block 10. This should include quantitative results (e.g., how often the toolkit correctly flags fake vs real in a test set of content) and qualitative analysis of failures. Also, include results of the adversarial tests – for example, show an instance where tweaking an image bypassed the watermark detector or where a deepfake passed as real, and explain why. The deliverable is both the report and any scripts/ notebooks used for the evaluation experiments.

Block 12 (Days 166–180): Capstone Project & Documentation​

  • Topics Covered: Final integration and knowledge synthesis. Students apply everything learned to a comprehensive capstone project: e.g., “Content Authenticity Guardian” – a full system that can be applied to a real-world scenario (such as verifying the authenticity of images circulating on social media during a breaking news event). Emphasis on documentation, usability, and reliability. This block also covers preparing thorough documentation and user guides, as well as a final review of the module’s topics.
  • Deliverables: Complete capstone project implementation of a content authenticity verification system, with a polished README or documentation site. A final presentation that demonstrates the system on real or realistic test cases (for instance, showing how a deepfake video call is flagged in real-time, or how an image with no provenance info but with an AI watermark is identified as AI-generated). The project package should include the code, sample data, and documentation. Additionally, a reference list of all research and resources used (creating a mini “Content Authenticity Handbook” for future reference, with citations of standards like C2PA, and notable research papers).

Full Tech Stack & Tools​

  • Programming & ML: Python (for scripting and ML prototyping), PyTorch/TensorFlow for deepfake detection models, OpenCV for video frame handling.
  • Datasets: Public deepfake datasets (FaceForensics++, Deepfake Detection Challenge dataset), image sets for watermark testing.
  • Watermarking: invisible-watermark Python library for image watermarking; OpenCV and PIL for image manipulations; pHash/PDQ hashing tools for fingerprinting.
  • Provenance & Signing: C2PA open-source SDKs (Rust or JavaScript) for creating and verifying content credentials; OpenSSL for managing keys/certificates for signing.
  • Real-Time Pipeline: GStreamer or FFmpeg for streaming pipeline integration; possibly PyQt/Streamlit for building a simple interface for the final toolkit.
  • DevOps: Docker (to encapsulate environment, especially for C2PA tools or ML models), and GPU access for deep learning tasks.
  • References & Standards: Academic papers on deepfake detection, industry articles on tools like FakeCatcher, official C2PA spec and documentation, and various GitHub repositories for deepfake detection and watermarking used throughout the program.