rigen-manandhar/truthlens-fake-news-detection

Purpose and Clients

TruthLens is a misinformation-risk assessment system, not a truth-verification engine. It combines language-model signals, source context, article extraction, claim hints, optional news coverage, and uncertainty handling. Primary clients are a Next.js web application and Chrome extension; both reach the same prediction boundary. Key entry points: Frontend/app/page.tsx, Frontend/app/fake-detection/page.tsx, extension/popup-main.js, Backend/app/main.py.

Web Application and API Boundary

The Next.js App Router application provides detection, deepfake detection, news, account, settings, and admin experiences. Browser-facing API routes act as a server-side boundary: prediction routes proxy/request backend analysis, while identity and product data remain handled in Next.js. Core paths: Frontend/app/api/predict/route.ts, Frontend/app/api/deepfake/predict/route.ts, Frontend/app/components/fakeDetection/useFakeDetectionController.ts.

Prediction Orchestration

FastAPI’s /predict flow is centralized in the prediction service. It accepts submitted text and/or URL, coordinates eligibility, extraction, source/evidence/model stages, and constructs a compatible structured response. This is the principal control-flow boundary between HTTP input and risk evaluation: Backend/app/main.py, Backend/app/services/predict_service.py, Backend/app/services/predict_response_builders.py.

Content Acquisition and Source Context

URLs are first screened for eligibility, then article text is safely extracted where possible; users may instead provide manual text. Source credibility is derived from a transparent local seed database rather than a live authority. Extraction provenance and source quality are inputs to final risk interpretation. Paths: Backend/app/url_eligibility.py, Backend/app/article_extractor.py, Backend/app/data/source_credibility.json.

Hybrid Language and Evidence Signals

Two local RoBERTa+LoRA artifacts contribute separate headline/claim and article-language signals. The hybrid layer runs and normalizes model outputs. Independently, the evidence layer identifies check-worthy claim hints and can query trusted-source coverage when NEWS_API_KEY is configured; missing API credentials explicitly leave coverage unchecked. Paths: Backend/app/hybrid_model.py, Backend/app/hybrid/model_runner.py, Backend/app/evidence.py; model assets: Backend/model/model_a/adapter_model.safetensors, Backend/model/model_b/adapter_model.safetensors.

Scoring, Explanation, and Uncertainty

Scoring fuses imperfect signals into LIKELY REAL, SUSPICIOUS, or UNCERTAIN, preserving the distinction between observed risk and factual truth. Confidence, model disagreement, conflicting evidence, extraction quality, limitations, and explanations are first-class response metadata rather than hidden implementation detail. Paths: Backend/app/scoring.py, Backend/app/explanations.py, Backend/app/schemas.py.

Deepfake Analysis

A distinct media-analysis pipeline is exposed separately from textual misinformation assessment. It handles media input, face detection, model loading/inference, and response schemas; the underlying Xception classifier code and training configuration are retained under the deepfake package. Paths: Backend/app/deepfake_detection/service.py, Backend/app/deepfake_detection/inference.py, Backend/deepfake/src/models/xception_classifier.py.

Identity, Persistence, and Operations

Next.js owns authentication, user/account lifecycle, privacy exports/deletions, sessions, feedback, and administration. It uses NextAuth-style auth components and a MongoDB client/server data layer, with email, audit, rate-limiting, and privacy services at the server boundary. Paths: Frontend/lib/auth.ts, Frontend/lib/mongodb-client.ts, Frontend/lib/server/privacy.ts, Frontend/app/api/auth/[...nextauth]/route.ts.