Now You See the Hate: Adaptive View Retrieval for Hidden Hateful Illusions

2026-07-23 · Source: cs.CV updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Cybersecurity & Data Privacy · Depth: Expert, extended

Summary

Adaptive View Retrieval (AVR) is a novel framework designed to detect hidden hateful optical illusions, addressing a critical gap where current multimodal safety systems achieve only 10.2% accuracy with state-of-the-art VLMs like GPT-4o and Gemini-2. AVR formulates this as a perceptual retrieval problem, employing a complementary view bank of image transformations (e.g., contrast, edge closure, figure-ground separation, low-pass smoothing). It adaptively selects the most informative views for each image, retrieves hidden message identities, and calibrates for harmfulness. Using a frozen CLIP encoder, AVR achieved 93.2% balanced accuracy on the HatefulIllusion test split, significantly surpassing original-view baselines and fixed single-transform filters. The system also matched or exceeded human performance on IllusionMNIST, IllusionFashionMNIST, and IllusionAnimals, and improved VLM hidden-content recognition on HC-Bench under the SemVink protocol.

Key takeaway

For AI Security Engineers or Machine Learning Engineers developing multimodal safety systems, you must move beyond original-view classification. Implement adaptive multi-view retrieval to expose hidden hateful content, as single-transform filters are insufficient and current VLMs fail significantly. Prioritize recovering hidden meaning before assessing harm to improve moderation accuracy and auditability. This is crucial for platforms managing libraries of known slurs and symbols.

Key insights

Effective multimodal moderation necessitates recovering hidden meaning prior to assessing its harmfulness.

Principles

• Hidden content detection is a perceptual retrieval problem.
• Adaptive view selection is crucial for diverse illusions.
• No single image transform suffices for all hidden patterns.

Method

Adaptive View Retrieval builds a complementary view bank from perceptual operations. It adaptively gates views, retrieves best-matching templates, and calibrates for harmfulness, separating identity recovery from moderation.

In practice

• Use multi-view template retrieval for hidden content.
• Apply perceptual transforms like low-pass and figure-ground.
• Integrate a calibration head for harmful/benign decisions.

Topics

• Adaptive View Retrieval
• Multimodal Safety
• Hateful Illusions
• Perceptual Retrieval
• Image Moderation
• CLIP Encoder

Best for: Computer Vision Engineer, Research Scientist, AI Scientist, Machine Learning Engineer, AI Security Engineer

Related on AIssential

• Now You See the Hate: Adaptive View Retrieval for Hidden Hateful Illusions — Robust multimodal moderation requires recovering hidden meaning before deciding whether it is harmful.
• HarmVideoBench: Benchmarking Harmful Video Understanding in Large Multimodal Models — Harmful video understanding in LVLMs requires multi-layered evaluation beyond binary classification and explicit rationales.
• The Synthetic Media Shift: Tracking the Rise, Virality, and Detectability of AI-Generated Multimodal Misinformation — AI-generated misinformation achieves high virality via passive engagement but is quickly corrected by community consensus.
• Beyond a Joke: Multi-Angle Reasoning for Detecting and Explaining Harmful Humor in Memes — MAR-12 uses multi-perspective VLM analysis to detect and explain harmful humor in complex memes.
• To Be Multimodal or Not to Be: Query-Adaptive Audio-Visual Person Retrieval via Active Modality Detection — Cross-modal score consistency reliably detects active modalities, preventing noise from absent modalities in multimodal retrieval.
• What the Eyes See, the LLMs Miss: Exploiting Human Perception for Adversarial Text Attacks — LLM content moderation systems are vulnerable to visually-based adversarial attacks that exploit human perception.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CV updates on arXiv.org.