On-device face and Indonesian e-KTP detection for Kotlin Multiplatform.

Faktel is an open-source Kotlin Multiplatform (KMP) library that runs computer-vision pipelines on the device - no cloud round trip, no vendor CV SDK. One shared Kotlin API for Android, iOS and desktop JVM.

Bahasa Indonesia

Status: pre-1.0 (0.x). The API can still change between minor versions; every change is recorded in the CHANGELOG. See Limitations before using it for anything security-sensitive.

All heavy math runs through one small seam, InferenceEngine, so the ML runtime is swappable.

Group ID: io.github.cybersafetyid.faktel. Details: architecture.

// build.gradle.kts of your shared module
kotlin {
    sourceSets {
        commonMain.dependencies {
            implementation("io.github.cybersafetyid.faktel:faktel:<version>")
        }
    }
}

Then wire the platform pieces (one-time): Android - iOS - Desktop JVM.

Shared code depends only on the InferenceEngine interface:

class SelfieChecker(engine: InferenceEngine, yunet: ByteArray, fas: ByteArray) {
    private val analyzer = FaceAnalyzer(
        detector = YuNetFaceDetector(engine, yunet),
        liveness = MiniFasNetLivenessDetector(engine, fas),
    )

    fun check(frame: RgbImage): FaceAnalysis = analyzer.analyze(frame)   // detect -> quality -> liveness
}

val scanner = KtpScanner(faceDetector = YuNetFaceDetector(engine, yunet))
val scan = scanner.scan(photo)
if (scan.isAcceptable) {
    val card = scan.card!!            // upright 1011x638 image
    val portrait = scan.portrait!!    // face box in card coordinates
} else {
    showHint(scan.issues)             // e.g. [BLURRY, GLARE] -> "hold steady, avoid reflections"
}

Selfie-to-KTP match: docs/guides/selfie-ktp-matching.md.

Start at docs/README.md. Highlights: Getting started - Architecture - Face guide - KTP guide - Models & licences - Tuning thresholds - Privacy & security - Troubleshooting - API reference.

• Liveness is passive and single-frame. It stops casual printed-photo and screen-replay attacks. It is not certified anti-spoofing (e.g. iBeta PAD) and must not be your only control in a high-assurance flow.
• The KTP module checks that an object looks like a KTP (shape, size, focus, glare, portrait placement). It does not read text (no OCR), verify authenticity, or detect forgeries.
• The bundled classical card detector needs a plain, contrasting background and a visible margin around the card. A learned detector is on the roadmap; the KtpDetector interface is where it plugs in.
• Default thresholds are starting points. Calibrate on your devices and users (guide).
• KTP images are personal data under Indonesian data-protection law. Process on-device, do not persist raw images (guide).

Unit tests run on JVM and the iOS simulator in CI. The real models (YuNet, MiniFASNetV2, an ArcFace MobileFaceNet) were exercised end-to-end through ONNX Runtime on desktop JVM and on an iPhone simulator with matching results. On-device Android runtime (as opposed to compilation) is exercised by your app; see docs/development/testing.md for what is and is not covered automatically.

Contributions are welcome - bug reports, docs, thresholds from real devices, new backends, a learned KTP detector. Read CONTRIBUTING.md and the Code of Conduct. Versioning follows SemVer. Security issues: SECURITY.md.

Apache License 2.0. Bundled model weights keep their own licences - see models/MODELS.md and NOTICE.