{ "lesson": "23-watermarking-synthid-stable-signature-c2pa", "title": "Watermarking - SynthID, Stable Signature, C2PA", "questions": [ { "stage": "pre", "question": "How does Kirchenbauer-style (SynthID-text) token watermarking embed a signal?", "options": [ "By encrypting the output with a public key", "At each decoding step, hash the previous K tokens to partition the vocabulary into green/red sets and bias sampling toward green; detection counts green tokens via a z-score", "By inserting a hidden HTML comment", "By modifying the prompt template" ], "correct": 1, "explanation": "" }, { "stage": "check", "question": "What is the primary limitation of text watermarks against paraphrase?", "options": [ "Paraphrase only affects image watermarks", "They actually survive paraphrase", "Paraphrase strengthens the watermark", "Rewording or paraphrasing the text destroys the green-token statistical signal" ], "correct": 3, "explanation": "" }, { "stage": "check", "question": "What does Stable Signature (Fernandez et al., ICCV 2023) do?", "options": [ "Adds a visible watermark in the corner of the image", "Signs the metadata with a private key", "Uses C2PA to sign the image at upload", "Fine-tunes the latent diffusion decoder so every generated image carries a fixed embedded message detectable from the latent" ], "correct": 3, "explanation": "" }, { "stage": "check", "question": "What does the May 2024 'Stable Signature is Unstable' result demonstrate?", "options": [ "The watermark is provably unforgeable", "The watermark requires C2PA to function", "The watermark fails only on JPEG compression", "Post-generation fine-tuning of the decoder can remove the watermark while preserving image quality, undermining adversarial robustness" ], "correct": 3, "explanation": "" }, { "stage": "post", "question": "How are C2PA metadata and watermarking complementary?", "options": [ "C2PA replaces watermarks entirely", "Metadata can be stripped but carries rich provenance chain; watermarks persist through transcoding but carry few bits; layered, they provide a usable provenance story", "Watermarks are signed by the same keys as C2PA", "They are interchangeable" ], "correct": 1, "explanation": "" }, { "stage": "post", "question": "Why is 'no SynthID signal' NOT proof of authenticity (human origin)?", "options": [ "Because SynthID is closed-source", "Because SynthID always produces false positives on human text", "Because SynthID only watermarks generations from SynthID-enabled models; content from any model without SynthID is unwatermarked by default", "Because SynthID only works on Gemini outputs" ], "correct": 2, "explanation": "" } ] }