Spaces:
Running
Running
| title: Prem | |
| emoji: 🔥 | |
| colorFrom: red | |
| colorTo: yellow | |
| sdk: static | |
| pinned: false | |
| short_description: In device co-pilot | |
| # Premithius V0.0.3 | |
| **India-First On-Device AI Security | Zero LLM Footprint | ARMadillo Architecture** | |
| Premithius is an uncompromising, edge-native security layer for Android devices. Designed specifically for the Indian cybersecurity context, it proactively intercepts sophisticated threats like ₹1 collect requests, Aadhaar/PAN data leaks, and zero-day scams before they ever reach the user's eyes—all without sending a single byte of personal data to the cloud. | |
| --- | |
| ## 🛡️ Kavacha Engine (5-Layer Defense) | |
| Inspired by NVIDIA's ARMadillo architecture, Premithius utilizes a DOCA FLOW-style early-exit pipeline. This ensures maximum efficiency, allowing standard traffic to pass cleanly while trapping anomalies with progressive computational weight. | |
| 1. **Layer 0 (Whitelist):** Immediate pass for trusted, verified senders. | |
| 2. **Layer 1 (India Threat Pack):** Hardened heuristic rules targeting specific local threat vectors (see below). | |
| 3. **Layer 2 (Vector Similarity):** Sub-millisecond KNN semantic search against offline scam embeddings using `sqlite-vec` (ARM NEON accelerated). | |
| 4. **Layer 3 (Random Forest):** A lightweight trained ensemble classifier exported to pure Dart via `m2cgen`. Trained on the HuggingFace `india-spam-sms` dataset with 13 custom numerical and categorical features. Scales dynamically during high-risk festival periods. | |
| 5. **Layer 4 (MobileBERT):** Our final deep-learning fallback layer. Integrates the `mrm8488/bert-tiny-finetuned-sms-spam-detection` model running locally via `tflite_flutter` and a custom-built native Dart WordPiece tokenizer. | |
| --- | |
| ## 🇮🇳 India Threat Pack | |
| Premithius is built for the ground reality of Indian digital fraud: | |
| * **UPI Firewall:** Natively catches the classic "₹1 collect request" marketplace fraud, scanning for suspicious VPA usage and urgency indicators. | |
| * **OTP Context Verifier:** Intercepts incoming OTP messages and cross-references them against the active foreground application to prevent OTP-forwarding phishing. | |
| * **Aadhaar/PAN Data Shield:** A final line of defense baked directly into the custom Premithius Keyboard, warning users *before* they inadvertently send sensitive documents to untrusted contacts. | |
| * **Festival Calendar:** Scam volume spikes ~340% during major holidays. The pipeline dynamically adjusts its sensitivity multiplier during these critical windows. | |
| --- | |
| ## 🚀 Edge-Native & Privacy First | |
| * **Zero LLM Overhead:** V0.0.3 maintains a tiny footprint by running Random Forest directly in Dart arithmetic and utilizing a highly optimized, quantized BERT module instead of a 900MB Qwen2 model. | |
| * **Data Shield Insights**: A custom Android native Kotlin telemetry layer continuously and securely queries `AppOpsManager` and `UsageStatsManager` in the background. It surfaces hardware access events (Mic/Camera) and compiles local privacy stories translating background behavior for users. | |
| * **Home Screen Widget**: A completely native Kotlin AppWidget provides users with 4x2 interactive real-time monitoring of their device's security status, avoiding Flutter engine wakelocks entirely. | |
| --- | |
| ## 🛠️ Build Requirements | |
| * Flutter SDK (^3.11.1) | |
| * Android Studio / Android SDK | |
| * *Note:* The final release build requires the `sqlite-vec-android:0.1.7` dependency. Ensure your Android `build.gradle` is configured with the correct Maven artifact repositories if retrieving `io.github.asg017` fails during compilation. | |
| ```bash | |
| # Debug Build | |
| flutter build apk --debug | |
| # Release Build | |
| flutter build apk --release | |
| ``` | |
| --- | |
| *Built for the next billion users. Because security shouldn't require surrendering your privacy.* | |