Deepfake Identity Verification Triage Pipeline with Google ADK and Contextual AI
As generative AI deepfakes increasingly target remote KYC onboarding and voice biometrics in digital banking, financial institutions require agentic validation frameworks capable of real-time multimodal verification. These systems must combine image artifact detection, liveness analysis, and Contextual AI contextual grounding to protect identity pipelines.
What you are building
The core problem, expected build, and operating context for this challenge.
Construct a multimodal identity verification pipeline using Google ADK and Contextual AI to evaluate biometric media artifacts and isolate synthetic impersonation attacks.
How work is evaluated
Validates the multimodal agent's ability to classify authentic vs deepfake video frames and verify document metadata against grounded knowledge bases.
Challenge datasets
Public datasets, plus the private datasets you have access to.
How submissions are scored
These dimensions define what the evaluator checks and which criteria separate a passable run from a strong one.
decision_enum_validity
decision_action must be ACCEPT, STEP_UP, or REJECT_AND_ESCALATE.
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
deepfake_detection_precision
Precision in identifying synthetic deepfake submissions. • target: 0.98 • range: 0.85-1
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
What you should walk away with
Build a multimodal agent using Google Agent Development Kit (ADK) and Gemini models
Integrate Contextual AI grounding tools to cross-reference identity documents with external registries
Implement deepfake artifact extraction for spectral and thermal video anomalies
Design zero-trust risk score aggregation for step-up step-down authentication routing
Reference links and supporting material
Dataset of 500 simulated KYC onboarding attempts containing high-resolution synthetic faces, deepfake swaps, and corresponding identity document metadata.
Challenge setup
Validates the multimodal agent's ability to classify authentic vs deepfake video frames and verify document metadata against grounded knowledge bases.
Challenge input
JSON containing payload with image_base64_hash, document_text, and claimed_identity.
Google ADK
Multimodal agent development SDK supported by Gemini models.
Contextual AI
Provides enterprise-grade grounding to prevent hallucinated KYC validation.
Evaluated output
JSON with is_authentic, deepfake_confidence, mismatch_reason, and decision_action.
- decision_action must be ACCEPT, STEP_UP, or REJECT_AND_ESCALATE.
- Precision in identifying synthetic deepfake submissions. • target: 0.98 • range: 0.85-1
- Benchmark: Deepfake Vision Evaluation Suite
- Deepfake Detection Precision target: 0.98
- 1 public reference case
- Python execution harness
- Python sandbox (unavailable on Versalist)
Summary shows the first task, up to three tools, and up to three evidence items. Open the technical recipe to view all configured items.
View technical recipe
Configured tools
- Google · Required
- Google ADK · Optional
- Google Jules · Optional
- Google Jules · Optional
Evaluation contract
- decision_enum_validity · Weight 1
- deepfake_detection_precision · Weight 1
Metrics
- deepfake_detection_precision · Target 0.98
Benchmarks
- Deepfake Vision Evaluation Suite
- security
Reference cases
- 1 public reference case
Workflow tasks
- Task 1
Input: JSON containing payload with image_base64_hash, document_text, and claimed_identity.
Output: JSON with is_authentic, deepfake_confidence, mismatch_reason, and decision_action.
Execution harness: Python
Evaluation tools
- Google ADK · Agent Architecture
- Contextual AI · Verification Grounding
Recipe freeze marker
No recipe freeze time is recorded.