FaceLivenessDetector Feature Request: Background environment lighting feedback

[sothon19931]

The Face Liveness Detection system produces high confidence scores for 3D masks, which are as confident as scores given to real users under bright lighting conditions. This raises a serious security concern, as spoofing using hyper-realistic masks can pass liveness checks. Meanwhile, real users in bright environments often receive low scores (e.g., ~78.2%) and fail the check due to overexposure or glare.

Observed on Android, but it's likely that the issue also exists on iOS.

Suggested improvements :

1. Strengthen spoof detection, especially against 3D masks.
2. Normalize lighting through preprocessing or auto-exposure control.
3. Guide users in real-time to reposition or adjust lighting when needed.

Found on : Latest version
com.amplifyframework.ui:authenticator v1.5.0
com.amplifyframework.ui:liveness v1.4.2

3D Mask :

Liveness Result of 3D Mask :

[harsh62]

Thanks for raising this issue. One of our team members will look into the issue and get back to you.

[sothon19931]

Thanks for responding. This issue is quite critical for our use case, and I’m looking forward to your findings.

[tylerjroach]

@sothon19931 Can you reach out to AWS Support so this can be looked at by our science teams (https://docs.aws.amazon.com/awssupport/latest/user/case-management.html). Any concerns with scores will need to be looked at by our science teams. The Android library is just the UI component.

[sothon19931]

@tylerjroach I understand that the background process for image and video analysis harnesses flash-based verification by analyzing colors reflected on the face during the freshness challenge on video streaming.
Apologies if I’ve misunderstood this.

Resulting in low confidence scores for real users in bright sunlight, could potentially be addressed by providing user guidance through the UI—for example, a static message suggestion similar to the demo in this repo, Or possibly by using CameraX ImageAnalysis.Analyzer to detect overexposure and provide real-time suggestions to adjust the user’s position or environment. This would help ensure that the video input reaching to the backgound process (AI,ML model) — is optimal.

However, Issue high confidence scores for hyper-realistic 3D masks that the underlying AI,ML model or related other backgound process may require further improvement to enhance anti-spoofing capabilities at a deeper level.

I’ll reach out to AWS Support so the science team can investigate this further.
If there's anything else you'd like to suggest, please feel free to let me know.
Thank you for your guidance!

[harsh62]

Resulting in low confidence scores for real users in bright sunlight, could potentially be addressed by providing user guidance through the UI—for example, a static message suggestion similar to the demo in this repo, Or possibly by using CameraX ImageAnalysis.Analyzer to detect overexposure and provide real-time suggestions to adjust the user’s position or environment. This would help ensure that the video input reaching to the backgound process (AI,ML model) — is optimal.

I'll mark this as a feature request.