How to add deepfake detection to your KYC flow
Document and liveness checks were designed before generative AI. Today, AI-generated selfies and synthetic ID images can slip through. Adding an AI-image detection step to onboarding catches likely-synthetic media before an account opens.
Where it fits in onboarding
- User uploads ID document + selfie.
- Your existing OCR / liveness / face-match runs.
- New step: send the ID image and selfie to an AI-image detector; get an AI-probability score per image.
- Route by risk: low → auto-approve; elevated → manual review or step-up verification.
Example: one API call
DEEF.AI exposes a simple endpoint authenticated by a license key (pay-per-use credits). Send the raw image bytes:
curl -X POST https://api.deef.ai/v1/detect \
-H "Authorization: Bearer $DEEF_KEY" \
--data-binary @selfie.jpg
# → {"verdict":"ai_generated","risk":0.94,"remaining":299}
Read risk (0–1) and set a threshold that matches your risk appetite. The full spec is at /openapi.json; the API catalog is at /.well-known/api-catalog.
Design tips
- Score, don't block outright. Use the probability to route to review — detectors aren't 100% accurate.
- Check both ID and selfie. Fraud often targets whichever you don't check.
- Keep an evidence trail. Store the verdict and (for disputes) the SHA-256-fingerprinted report.
- Privacy. Only the deep-scan API uploads bytes for inference; images aren't retained after the result.
FAQ
Why does KYC need AI-image detection?
Synthetic IDs and AI selfies can defeat pre-genAI checks; detection flags them before onboarding completes.
How hard is integration?
One POST per image, authenticated with a license key. Read the risk score and route accordingly.
DEEF.AI provides screening-grade decision support. No detector is 100% accurate; use scores as one signal in your KYC risk model.
KYC 流程如何接入 deepfake / AI 图像检测
证件与活体检测是在生成式 AI 之前设计的。如今 AI 生成的自拍和合成证件图可能蒙混过关。在开户流程里加一步 AI 图像检测,能在开户前拦下大概率合成的素材。
它在开户流程中的位置
- 用户上传证件 + 自拍。
- 你现有的 OCR / 活体 / 人脸比对照常运行。
- 新增一步:把证件图和自拍发给 AI 图像检测器,得到每张图的 AI 概率分。
- 按风险分流:低 → 自动通过;偏高 → 人工复核或加强验证。
示例:一次 API 调用
DEEF.AI 提供一个用卡密鉴权的简单端点(按次 credits)。直接发原始图像字节:
curl -X POST https://api.deef.ai/v1/detect \
-H "Authorization: Bearer $DEEF_KEY" \
--data-binary @selfie.jpg
# → {"verdict":"ai_generated","risk":0.94,"remaining":299}
读取 risk(0–1),设一个匹配你风险偏好的阈值。完整规范见 /openapi.json,API 目录见 /.well-known/api-catalog。
设计建议
- 用分数分流,别直接拦死。用概率路由到复核——检测器并非 100% 准确。
- 证件和自拍都查。欺诈往往挑你没查的那张下手。
- 留证据链。保存判定结果,争议时保留含 SHA-256 指纹的报告。
- 隐私。只有深扫 API 会上传字节做推理;返回结果后不保留图像。
常见问题
KYC 为什么需要 AI 图像检测?
合成证件与 AI 自拍能骗过生成式 AI 之前的检查;检测能在开户完成前标记它们。
集成难吗?
每张图一次 POST,用卡密鉴权。读风险分并据此分流即可。
DEEF.AI 提供初筛级决策辅助。没有任何检测器能做到 100% 准确;请把分数作为 KYC 风险模型里的一个信号。