AIdeas Finalist: LikenessGuard - Multi-Agent, Edge-Capable Consent Enforcement Platform with Cross-Platform Integration and Cryptographic Proof-of-Face
LikenessGuard enforces consent before any AI generates or edits an image involving a real person’s likeness. Built with multi-agent Bedrock, edge enforcement, Proof-of-Face, and live integrations with Claude and Grok, it prevents harm at the source and will be open-sourced after April 30, 2026, so AI platforms, developers, researchers, and regulators worldwide can adopt, extend, and improve it freely.
App Category:
social-impact
What to Expect from This Article
This article follows a structured narrative designed for clarity and depth:
- My Vision → What the idea is and what was built in v2
- Why This Matters → The real-world problem of non-consensual AI imagery and why prevention is critical
- Demo → Visual demonstration of the working prototype and live cross-platform examples
- How I Built This → Technical approach, full architecture, key v2 improvements, and cross-platform integrations (Claude and Grok)
- What I Learned → Direct response to judge feedback and key insights gained during development
- Future Vision → Open-source plans and long-term industry adoption as a responsible AI framework
MY VISION
LikenessGuard is the world’s first proactive, multi-agent consent enforcement platform that shifts AI image generation from “generate-first, moderate-later” to mandatory pre-generation verification.
Users register their likeness once (or benefit from proactive default-deny protection), define granular machine-readable policies, and LikenessGuard blocks unauthorized use at the source before any AI model (OpenAI, Claude, Grok, etc.) generates or edits the image.
What I built in v2
- Full multi-agent Bedrock system (Anomaly & Threat Agent, Policy Reasoner, Consent Orchestrator)
- Hybrid facial matching with Rekognition + Titan Embeddings + OpenSearch vector search
- Edge enforcement with AWS IoT Greengrass (offline default-deny)
- Cryptographic Proof-of-Face using AWS KMS + C2PA-compatible manifests
- Federated registry API with signed peer requests
- Cross-platform integrations: Claude via MCP connector and Grok via function calling
- Intelligent dashboard with rich audit viewer, live metrics, and agent reasoning traces
The system now works seamlessly across platforms while maintaining a unified audit trail in the LikenessGuard dashboard.
Why This Matters
Non-consensual AI-generated imagery continues to harm millions - women, minors, public figures, and everyday people. Current moderation happens too late: the damage is already done.
LikenessGuard inverts this model. By enforcing consent before generation, it prevents harm at the source, restores trust in AI, reduces platform moderation costs, and protects vulnerable populations through default-deny and edge capabilities.
It aligns directly with the EU AI Act, C2PA provenance standards, and global privacy regulations, offering a technical foundation for responsible generative AI.
Deepfakes are no longer a future threat they are costing millions and destroying lives today. Just look at what’s already happened:
1: A finance employee at Arup was deceived by a deepfake video call posing as company executives, leading to 15 fraudulent transfers totaling HK$200 million (~$25.6M) in a single day.
2:A Hong Kong-based group of 27 suspects used AI generated female personas (video, images, and voice) on dating apps to build trust, then scammed victims into fake crypto investments. The operation targeted people across multiple countries and caused large-scale emotional and financial damage before being shut down by police.
3: AI generated explicit images (January 2024) of Taylor Swift spread rapidly on platforms like X and Telegram, with one image reaching 47 million views before removal. The incident exposed the growing scale of deepfake porn (over 90% targeting women), triggering global backlash, responses from the White House, and renewed efforts to pass laws against non-consensual AI generated intimate content.
4: Hackers compromised a Ukrainian news website and TV channel ticker (March 2022) to broadcast a deepfake video of President Volodymyr Zelensky urging troops to lay down arms and surrender to Russia. The ~1-minute clip (poor lip-sync but initially convincing to some) spread rapidly on social media amid the invasion. Zelensky quickly debunked it in a real video
5:Hundreds of global cases (Ongoing, Spiking 2025–2026) have involved students’ faces being taken from social media or school photos and placed onto explicit AI generated images. Incidents across the US, Japan, India, and Europe ranging from high school victims to over 160 elementary-age children in Japan alone have led to bullying, severe emotional trauma, harassment, and in some cases links to suicide, highlighting a growing crisis of non-consensual deepfake abuse among youth.
Demo
The video demonstrates the end-to-end flow: prompt without and with protection, LikenessGuard multi-agent consent check, cross-platform examples (Claude and Grok), and Proof-of-Face visibility in the dashboard.
How I Built This
I built LikenessGuard entirely on AWS using a serverless-first architecture, heavily leveraging Kiro for rapid development and refinement.The core is a Supervisor Lambda that orchestrates a 9-step consent pipeline in under 300ms:
- Input validation — accepts base64 images or public URLs
- Face detection — Amazon Rekognition confirms a human face is present
- Facial embedding — Bedrock Titan Embed Image v1 generates a 512-dimensional vector
- Vector matching — OpenSearch Serverless k-NN finds the top-10 most similar registered faces
- Anomaly screening — Claude Haiku checks for prompt injection, jailbreaks, and rate abuse
- Policy evaluation — Nova Pro reasons over the subject's consent policy and the usage type
- Cryptographic signing — KMS ECDSA P-256 signs a C2PA-compatible Proof-of-Face manifest
- Audit logging — DynamoDB records an immutable, tamper-evident audit entry
- Response assembly — Returns decision, reasoning trace, and signed proof
The Multi-Agent System
Three Bedrock agents collaborate on every consent decision:
The Anomaly Agent (Claude Haiku) runs first as a fast-path gate. It checks a DynamoDB suspension list and scans for injection patterns in under 20ms. The LLM is only invoked for borderline cases, keeping costs low.
The Consent Orchestrator (Nova Pro, with Claude Sonnet as fallback) is the primary decision-maker. It evaluates the subject's policy against the requested usage type, applying the most-restrictive interpretation when policy clauses are ambiguous. It returns a structured decision with confidence score and reasoning trace.
The Policy Reasoner (Nova Lite) handles natural language policy conversion. Subjects can write policies in plain English — "allow editorial use but deny face swaps and commercial use" and the reasoner converts them to structured JSON with conflict detection.
Hybrid Facial Matching
v1 used SHA-256 fingerprint hashing, which had a 1.2% false negative rate. v2 replaced this with a hybrid pipeline:
- Rekognition detects the face and extracts the bounding box
- Titan Embed Image generates a 512-dimensional semantic vector
- OpenSearch Serverless runs HNSW k-NN search (cosine similarity) returning the top-10 candidates in ~25ms
- The Consent Orchestrator applies confidence scoring on top of the similarity score
This reduced false negatives to below 0.5% and scales to millions of registered faces without performance degradation.
Proof-of-Face (Cryptographic)
Every ALLOW decision produces a KMS-signed C2PA-compatible manifest containing the subject ID, requester, decision, similarity score, and a JWKS-verifiable ECDSA signature. If KMS signing fails, the decision flips to DENY - the system never returns an unsigned ALLOW. A public JWKS endpoint on CloudFront enables any third party to verify the proof independently.
Edge Enforcement
An AWS IoT Greengrass v2 component runs on edge devices with a SQLite cache of facial vectors. When offline, it performs local cosine similarity matching and defaults to DENY when no cache match exists. On reconnect, it syncs offline decisions back to the central audit log.
Cross-Platform AI Integration
The most significant v2 addition: LikenessGuard now integrates directly with AI assistants so they check consent automatically.
Claude (Anthropic): I built an MCP (Model Context Protocol) server that exposes LikenessGuard tools -
check_consent , upload_image_for_consent , get_policy , update_policy . Connected via Claude.ai's connector system, Claude automatically calls LikenessGuard before generating or editing any image involving a real person. When consent is denied, Claude refuses the request and explains why.Grok (xAI): Using xAI's OpenAI-compatible API with function calling, Grok calls
check_consent before any likeness-related generation. The integration script uses the same API endpoint, producing audit entries visible in the same dashboard alongside Claude's requests.Both platforms produce unified audit logs the dashboard shows requester names (claude-ai-user, grok-xai) alongside every consent decision.
Performance
Metric | Target | Achieved |
|---|---|---|
P95 end-to-end latency | < 300ms | ~263ms |
False negative rate | < 0.5% | < 0.5% |
OpenSearch k-NN query | < 30ms | ~25ms |
KMS signing | < 20ms | ~18ms |
Anomaly Agent fast-path | < 20ms | ~15ms |
Cost at Scale
Service | 100k checks/month |
|---|---|
Lambda (Supervisor + agents) | $1.92 |
Bedrock (Nova Pro + Haiku + Titan) | $1.90 |
Rekognition DetectFaces | $1.00 |
KMS signing | $0.30 |
DynamoDB | $0.13 |
OpenSearch Serverless | Free Tier |
Total | ~$4.20/month |
The entire platform runs for under $5/month at 100,000 consent checks - less than a cup of coffee per day.
Technical Approach & Architecture
The architecture consists of:
- API Gateway + Lambda for all endpoints
- OpenSearch Serverless for scalable vector similarity search (Titan embeddings)
- Bedrock multi-agent orchestration (Claude Haiku, Nova Pro, Nova Lite)
- AWS IoT Greengrass for edge/offline consent checks with default-deny
- AWS KMS for Proof-of-Face cryptographic signing
- DynamoDB for policies and audit logs
- S3 for temporary image storage
Key v2 Improvements
(directly addressing judge feedback):
- Hybrid matching reduced false negatives to <0.5%
- Proactive discovery + public opt-out protects unregistered users
- Federated registry with JWT + KMS-signed peer requests
- Full Proof-of-Face MVP with verifiable C2PA manifests
- Edge enforcement on laptop (simulated production device)
Cross-Platform Integrations:
- Claude.ai: MCP server with streamable HTTP transport and OAuth handshake
- Grok (xAI): OpenAI-compatible function calling wrapper
What I Learned
Judge feedback was invaluable and directly shaped v2.
- Active registration limitation: Originally, protection required users to actively register their likeness. In v2, this was solved by introducing proactive discovery (automatically detecting similar faces and creating provisional DENY-ALL entries), a public opt-out endpoint that allows anyone to register a DENY-ALL policy without an account, and edge-based default-deny that protects unregistered or offline users automatically. These changes ensure vulnerable populations, minors, and non-active users receive meaningful protection without needing to proactively register.
- Similarity threshold and scale: Originally, the system relied on a fixed 95% similarity threshold with DynamoDB full-table scans, resulting in a 1.2% false negative rate and poor scalability. In v2, this was completely replaced with Amazon OpenSearch Serverless using Titan embeddings for high-dimensional vector search (HNSW index) combined with Nova Pro confidence scoring. This hybrid approach reduced false negatives to under 0.5%, enabled sub-30ms lookups regardless of registry size, and provided much more reliable matching across different angles, lighting conditions, and expressions.
- Platform adoption and federation: Originally, LikenessGuard was a standalone service that depended on individual platforms voluntarily integrating it, which limited its real-world effectiveness. In v2, this was addressed by implementing a lightweight federated registry API with JWT authentication and KMS-signed peer requests. This allows different AI platforms to securely share consent decisions and likeness data. Python and Node.js SDKs were also created, along with ready-to-use integration examples for Stable Diffusion, DALL·E, and ComfyUI, reducing integration time to 1–2 days and making adoption significantly easier for both platforms and developers.
- Proof-of-Face conceptual: Originally, Proof-of-Face existed only as a conceptual idea in the design. In v2, it was fully implemented as a working MVP using AWS KMS (ECDSA P-256 signing) to create C2PA 1.3-compatible signed manifests. Every ALLOW decision now generates a verifiable cryptographic signature attached to the output. A dedicated verification endpoint (/v2/proof/verify) was added so platforms or auditors can independently validate the consent record at any time, turning Proof-of-Face from a future vision into a production-ready trust layer.
Key insights from development:
Default-deny is the only safe default. Every error path in LikenessGuard returns DENY. KMS signing fails? DENY. Face not detected? DENY. Offline with no cache? DENY. OpenSearch unavailable? Fall back to DynamoDB, and if that fails too, DENY. This principle - never return an unsafe ALLOW -is the single most important design decision in the entire system.
Serverless keeps costs honest. The entire platform costs $4.20/month at 100k checks. No EC2 instances running idle. No reserved capacity. Every dollar spent is directly tied to a consent check that protected someone's likeness. That cost structure makes it viable for any platform to adopt, regardless of size.
Ethics must be encoded in system defaults, not added later. Kiro accelerated iteration of multi-agent workflows and dashboard components. Transparent audit logs and agent reasoning traces build user trust far more effectively than black-box moderation.
Future Vision
After the 10,000 AIdeas competition concludes on April 30, 2026, the complete LikenessGuard v2 codebase - including the multi-agent system, edge component, Proof-of-Face implementation, federated registry, Claude MCP connector, and Grok integration - will be released as open source under the Apache 2.0 license.
My clear intention is that AI platforms can adopt LikenessGuard internally as a pre-generation consent layer, government and regulatory bodies can reference or adopt this as a technical framework under responsible AI guidelines and enforce it where appropriate, the live API can be used directly as it is already running on AWS, or anyone can deploy the full system in their own AWS accounts once the code is open-sourced.
The long-term vision is for LikenessGuard to become the open, verifiable global standard for responsible AI consent enforcement so that platforms, developers, researchers, and regulators worldwide can adopt, extend, and improve it freely.
End of Article
This article was created for the AWS 10,000 AIdeas competition to demonstrate practical implementation of consent enforcement in AI systems. LikenessGuard v2 is fully functional and deployed on AWS infrastructure.
Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article