U-EARDNet: A New Adversarially Robust Solution for Multimodal Toxic Content Detection

This article introduces U-EARDNet, an end-to-end multimodal deep learning model that integrates text and visual features through an innovative gated fusion mechanism. It aims to effectively detect online toxic content and resist adversarial attacks. The model balances accuracy and adversarial robustness, providing a new technical solution for the field of content security.

The development of social media and content platforms has brought severe issues with online toxic content (such as malicious comments, hate speech, cyberbullying, etc.). Its governance faces three major challenges:

1. Multimodal nature: Toxic content often combines text, images, and other elements, making single-modal detection insufficient.
2. Adversarial attacks: Malicious users bypass detection through methods like homophone replacement and image perturbation.
3. Context dependency: Word meanings change with context, leading to frequent misjudgments in keyword matching.

The core innovation of U-EARDNet is its gated multimodal fusion mechanism, which dynamically adjusts the fusion weights of text and visual features (traditional methods mostly use concatenation or weighted average). Its architecture includes three components:

1. Text encoder: Based on a Transformer pre-trained model, it captures semantic, emotional, and toxic signals.
2. Visual encoder: Uses CNN or Vision Transformer to extract image features (including meme text and visual elements).
3. Gated fusion module: Through a learnable gate network, it combines statistical information of text/visual features and cross-modal attention scores to output normalized weights for feature fusion.

Adversarial attack threats include:

• Text level: Homophone replacement, special symbol insertion, etc.
• Image level: Tiny noise, color changes, etc.
• Cross-modal level: Text and image are normal individually but toxic when combined. U-EARDNet's defense strategies:

1. Adversarial training: Incorporate adversarially perturbed samples into training.
2. Feature space regularization: Make features insensitive to small perturbations.
3. Multi-scale feature fusion: Extract and fuse features from local to global granularities to increase attack difficulty.

The model was evaluated on multiple datasets: Hateful Memes (toxic memes), Toxic Comment Classification (toxic comments), and a custom adversarial test set. Evaluation metrics include accuracy, precision, recall, F1 score, and adversarial robustness indicators. Results show:

• U-EARDNet leads in accuracy on standard test sets.
• Its performance degradation under adversarial attacks is smaller than baseline methods.
• It has strong cross-modal understanding capabilities.
• The F1 score on the adversarial test set is 15-20 percentage points higher than traditional methods.

Application scenarios:

• Social media platforms: Real-time detection of text-image mixed content.
• Online communities: Forums, comment sections, bullet screens (configurable sensitivity thresholds). Deployment optimizations:

1. Model quantization: Low-precision weights reduce resource consumption.
2. Knowledge distillation: Lightweight student models lower inference costs.
3. Batch inference: GPU parallelism improves throughput.

Current limitations:

1. Language coverage: Mainly supports English, with limited support for other languages.
2. Emerging attacks: Requires retraining to handle new types of attacks.
3. Computational cost: Increased by multimodal fusion and adversarial training. Future directions:

• Expand to more modalities (audio, video).
• Continuous learning to adapt to the evolution of attacks.
• Enhance interpretability (visualize decision-making basis).

U-EARDNet is an important advancement in the field of multimodal toxic content detection, achieving a balance between accuracy and robustness through gated fusion and adversarial robustness design. It provides references for researchers and engineers in content security, AI ethics, and social media governance. Its technical ideas can be transferred to other AI security applications. Open-source code lays the foundation for community collaborative improvement and promotes technological progress in the field.