MiniFAS selection uses the MobileNetV3 inference path

[phucvinh57]

Discussed in #144

^{Originally posted by Tunahanyrd July 13, 2026}
I traced the anti-spoof model selection in 1.4.1 and found that the selected MiniFAS model type is lost before inference.

• minifas-v2 is registered as a selectable anti-spoof model:
  

  biopass/app/src-tauri/src/db.rs

  Lines 48 to 59 in aa8c3e2

  	KnownModel {
  	slug: "mobilenetv3-antispoof",
  	filename: "mobilenetv3_antispoof.onnx",
  	display_name: "MobileNetV3 Anti-Spoof",
  	model_type: "anti_spoofing",
  	},
  	KnownModel {
  	slug: "minifas-v2",
  	filename: "minifas_v2.onnx",
  	display_name: "MiniFASNet V2",
  	model_type: "anti_spoofing",
  	},

• The inference caller resolves the selected model path, but constructs FaceAntiSpoofing with only the path, image size and threshold:
  

  biopass/auth/face/antispoofing/antispoof_check.cc

  Lines 30 to 42 in aa8c3e2

  	bool checkAntiSpoofByAIModel(const FaceMethodConfig& faceCfg, const std::string& username,
  	const ImageRGB& face, const AuthConfig& authCfg,
  	const ModelRegistry& model_registry) {
  	const std::string modelPath =
  	model_registry.resolveModelPath(faceCfg.anti_spoofing.model.model_id).value_or("");
  	if (modelPath.empty() || !std::ifstream(modelPath).good()) {
  	spdlog::error("FaceAuth: Anti-spoofing model file not found: {}", modelPath);
  	return false;
  	}
  	try {
  	FaceAntiSpoofing face_as(modelPath, 128, faceCfg.anti_spoofing.model.threshold);
  	const SpoofResult result = face_as.inference(face);

• The omitted fourth argument defaults to mobilenetv3:
  

  biopass/auth/face/antispoofing/face_as.h

  Lines 19 to 23 in aa8c3e2

  	// model_type: "minifasv2" or "mobilenetv3" (default)
  	class FaceAntiSpoofing {
  	public:
  	FaceAntiSpoofing(const std::string& ckpt, int imgsz = 128, const float threshold = 0.8,
  	const std::string& model_type = "mobilenetv3");

• Filename inference only runs when model_type is neither mobilenetv3 nor minifasv2. Since the default is already mobilenetv3, the MiniFAS filename is never inspected:
  

  biopass/auth/face/antispoofing/face_as.cc

  Lines 24 to 34 in aa8c3e2

  	FaceAntiSpoofing::FaceAntiSpoofing(const std::string& ckpt, int imgsz, const float threshold,
  	const std::string& model_type)
  	: threshold(threshold),
  	imgsz(imgsz),
  	model_type(model_type),
  	session(ckpt, "FaceAntiSpoofing") {
  	// Unrecognized model_type: infer from the checkpoint filename.
  	if (this->model_type != "minifasv2" && this->model_type != "mobilenetv3") {
  	this->model_type =
  	(ckpt.find("mobilenetv3") != std::string::npos) ? "mobilenetv3" : "minifasv2";
  	}

The loaded file can therefore be minifas_v2.onnx while preprocessing and output handling still follow the MobileNetV3 branch.

I would like to send a small PR that passes the selected built-in model type explicitly and adds a regression test. The scoring policy and other anti-spoof behavior would remain unchanged.