Face spoof detection

Detect presentation attacks using anti-spoofing models

FaceNet Android includes face anti-spoofing (liveness detection) to identify when someone attempts to fool the system with a photo, video, or 3D model instead of a real face.

How it works

The spoof detector uses MiniFASNet models from Silent-Face-Anti-Spoofing. It analyzes the face at two different scales to detect presentation attacks:

1. Crops the detected face at 2.7x scale → Model 1
2. Crops the same face at 4.0x scale → Model 2
3. Combines predictions using softmax averaging
4. Returns spoof/real classification and confidence score

The multi-scale approach helps detect:

• Printed photos
• Digital screens (phones, tablets)
• Video replays
• 3D masks (with reduced accuracy)

The models work by analyzing texture patterns and Fourier transform features that differ between real faces and presentation attacks.

Implementation

The spoof detector is implemented in FaceSpoofDetector.kt:36-89 as a singleton:

FaceSpoofDetector.kt

@Single
class FaceSpoofDetector(
    context: Context,
    useGpu: Boolean = false,
    useXNNPack: Boolean = false,
    useNNAPI: Boolean = false,
) {
    data class FaceSpoofResult(
        val isSpoof: Boolean,
        val score: Float,
        val timeMillis: Long,
    )

    private val scale1 = 2.7f
    private val scale2 = 4.0f
    private val inputImageDim = 80
    private val outputDim = 3

    private var firstModelInterpreter: Interpreter
    private var secondModelInterpreter: Interpreter
    
    init {
        val interpreterOptions = Interpreter.Options().apply {
            if (useGpu) {
                if (CompatibilityList().isDelegateSupportedOnThisDevice) {
                    addDelegate(GpuDelegate(CompatibilityList().bestOptionsForThisDevice))
                }
            } else {
                numThreads = 4
            }
            useXNNPACK = useXNNPack
            this.useNNAPI = useNNAPI
        }
        firstModelInterpreter = Interpreter(
            FileUtil.loadMappedFile(context, "spoof_model_scale_2_7.tflite"),
            interpreterOptions,
        )
        secondModelInterpreter = Interpreter(
            FileUtil.loadMappedFile(context, "spoof_model_scale_4_0.tflite"),
            interpreterOptions,
        )
    }
}

Using the detector

The detector processes face bounding boxes from the camera frame:

FaceSpoofDetector.kt

suspend fun detectSpoof(
    frameImage: Bitmap,
    faceRect: Rect,
): FaceSpoofResult

Example usage

val spoofDetector = get<FaceSpoofDetector>()
val result = spoofDetector.detectSpoof(
    frameImage = cameraBitmap,
    faceRect = detectedFaceRect
)

if (result.isSpoof) {
    println("Spoof detected! Score: ${result.score}")
    println("Detection took ${result.timeMillis}ms")
} else {
    println("Real face detected. Score: ${result.score}")
}

Preprocessing details

The detector performs specific preprocessing before model inference:

1. Multi-scale cropping

FaceSpoofDetector.kt

val croppedImage1 = crop(
    origImage = frameImage,
    bbox = faceRect,
    bboxScale = scale1,  // 2.7x
    targetWidth = inputImageDim,
    targetHeight = inputImageDim,
)

val croppedImage2 = crop(
    origImage = frameImage,
    bbox = faceRect,
    bboxScale = scale2,  // 4.0x
    targetWidth = inputImageDim,
    targetHeight = inputImageDim,
)

2. RGB to BGR conversion

The models expect BGR color format:

FaceSpoofDetector.kt

for (i in 0 until croppedImage1.width) {
    for (j in 0 until croppedImage1.height) {
        croppedImage1[i, j] = Color.rgb(
            Color.blue(croppedImage1[i, j]),
            Color.green(croppedImage1[i, j]),
            Color.red(croppedImage1[i, j]),
        )
    }
}

3. Softmax averaging

Predictions from both scales are combined:

FaceSpoofDetector.kt

val output = softMax(output1[0])
    .zip(softMax(output2[0]))
    .map { (it.first + it.second) }
val label = output.indexOf(output.max())
val isSpoof = label != 1  // Label 1 = real face
val score = output[label] / 2f

Performance tuning

By default, the detector uses CPU with 4 threads:

val spoofDetector = FaceSpoofDetector(
    context = context,
    useGpu = false,      // CPU execution
    useXNNPack = false,  // Disable XNNPACK optimization
    useNNAPI = false,    // Disable NNAPI
)

Enabling GPU acceleration

GPU acceleration may not improve performance for these small models and can cause compatibility issues on some devices.

val spoofDetector = FaceSpoofDetector(
    context = context,
    useGpu = true,       // Enable GPU delegate
    useXNNPack = false,
    useNNAPI = false,
)

Using NNAPI

val spoofDetector = FaceSpoofDetector(
    context = context,
    useGpu = false,
    useXNNPack = true,   // XNNPACK for CPU optimization
    useNNAPI = true,     // Android NNAPI
)

Performance metrics

Typical inference times on mid-range devices:

Configuration	Inference Time	Total Time (with preprocessing)
CPU (4 threads)	~15-25ms	~30-45ms
GPU	~20-30ms	~35-50ms
NNAPI	~10-20ms	~25-40ms

The detector measures and returns timing in the FaceSpoofResult.timeMillis field, which appears in the app’s performance metrics.

Accuracy considerations

The anti-spoofing models have limitations:

Effective against:

• Printed photos (90%+ accuracy)
• Digital screens (85%+ accuracy)
• Simple video replays (80%+ accuracy)

Less effective against:

• High-quality 3D masks (50-70% accuracy)
• Advanced video replay attacks
• Adversarial attacks

Anti-spoofing is not foolproof. For high-security applications, combine it with other verification methods like challenge-response or depth sensors.

Model sources

The TFLite models were converted from PyTorch weights:

1. Original PyTorch weights from deepface
2. Converted via ONNX to TensorFlow
3. Quantized to FP16 for mobile deployment

See the conversion notebook in the source repository.