Visual-Based Spam Filter Achieves Over 98% Detection Accuracy

Global: Visual-Based Spam Filter Achieves Over 98% Detection Accuracy

A new visual-based spam detection system has been introduced by researchers, aiming to counter sophisticated email obfuscation methods. The study, posted to arXiv in December 2025, describes a multi‑step architecture that mimics human visual processing to identify spam that evades traditional text‑only filters.

Background on Visual Spam Techniques

Recent spam campaigns have increasingly employed visual tricks such as poisoned text, word obfuscation, and hidden‑text salting to bypass conventional detection engines. These tactics manipulate how characters appear on screen, making it difficult for text‑based classifiers to recognize malicious content.

Architecture of the VBSF System

The proposed Visual‑Based Spam Filter (VBSF) renders incoming emails exactly as they would appear to a user, capturing both the rendered text and the visual layout. This dual representation forms the foundation for two parallel analysis pipelines.

Dual Processing Pipelines

The first pipeline extracts rendered text via optical character recognition (OCR) and subjects it to naive Bayes and decision‑tree classifiers. Simultaneously, the second pipeline feeds the screenshot of the email into a custom convolutional neural network that evaluates visual patterns indicative of spam.

Meta‑Classifier Integration

Outputs from the text‑ and image‑based classifiers are combined using a stacking ensemble meta‑classifier. This approach leverages the strengths of each model, improving overall decision robustness.

Performance Evaluation

Testing on a purpose‑built dataset showed that VBSF achieved an accuracy exceeding 98%, surpassing existing spam detection techniques evaluated under the same conditions.

Implications for Email Security

The results suggest that incorporating visual analysis can substantially strengthen defenses against evolving spam tactics, offering a complementary layer to traditional text‑centric solutions.

This report is based on information from arXiv, licensed under Academic Preprint / Open Access. Based on the abstract of the research paper. Full text available via ArXiv.