Learning Strategies for Sensitive Content Detection: A Comprehensive Review

Existing surveys on sensitive content detection often have limitations. Some, like [3], focus on CSAM detection with an emphasis on legal and political aspects, but their coverage of deep learning techniques, particularly modern architectures like Vision Transformers (ViTs), is outdated. Other reviews, such as Pour et al. [4], concentrate on video content rating (VCR) systems and deep learning, neglecting traditional methods like text-based analysis (hashes, metadata) and older techniques such as skin color segmentation. Cifuentes et al. [5] reviewed explicit video detection up to 2019-2020, primarily focusing on deep learning and visual features, while overlooking textual and audio-only methods, as well as recent advancements like visual attention mechanisms. This survey distinguishes itself by offering a comprehensive and systematic study that consolidates research contributions across various content-based strategies. It uniquely brings together textual (hashes, keywords) and visual/auditory features, classifying methodologies to facilitate comparison and identify research gaps. This work aims to serve as a foundational resource for new researchers and guide future explorations in sensitive content detection.

Text-based strategies for sensitive content detection do not rely on the visual or auditory content itself but rather on associated textual information. These methods include image hash databases, web crawlers, and filename/metadata analysis. Image hash databases, such as Microsoft's PhotoDNA, generate unique digital signatures for images, allowing for the identification of known illegal content by comparing hashes against a database of previously reported material. While effective for detecting duplicates of existing illegal images, this method cannot identify novel CSAM. Web crawlers, like Project Arachnid, systematically scan web pages, collect content, and use technologies like PhotoDNA for hashing. These systems often integrate hash lists from law enforcement agencies (e.g., NCMEC, RCMP, Interpol) for CSAM detection. Filename and metadata analysis involves extracting information from file names and associated metadata to classify content. Conventional machine learning classifiers like Support Vector Machines (SVM) and logistic regression are often employed for this purpose. Commercial tools also exist, utilizing whitelists and blacklists based on metadata. However, these text-based methods can be inefficient, as sensitive content can be disguised with irrelevant text, rendering simple label-based analysis insufficient.

To overcome the limitations of single-modality analysis, strategies incorporating motion, audio, and multimodal features have been developed for sensitive content detection. Videos inherently possess spatiotemporal data that can provide richer information than static frames alone. Multimodal approaches, combining visual and auditory features, have demonstrated superior accuracy [25-27]. For instance, late fusion processes can combine fragment classifiers to identify sensitive scenes by leveraging various video data aspects like motionless frames and audio streams [27]. These methods have shown higher true positive rates and lower false positive rates compared to other techniques. However, challenges remain, such as detecting clothed individuals performing sexual actions with static movements or identifying subtle audio cues like moaning. Audio detectors, often based on low-level features, can also increase false positives when not combined with high-level features derived from deep learning. Accurate audio-based detection requires analyzing not only timbre but also chroma, amplitude, and other elements. Consequently, these multimodal strategies are increasingly integrated with successful deep learning solutions.

The performance of sensitive content detection strategies is heavily influenced by the heterogeneity of datasets, the effectiveness of feature extraction techniques (hashes, visual, audio), and the sophistication of the learning algorithms employed. While deep learning models have demonstrated superior performance, challenges persist. These include the detection of nuanced or novel forms of sensitive content, the potential for adversarial attacks to bypass detection systems, and the ethical considerations surrounding data privacy and bias in algorithms. Future research directions lie in developing more robust and generalizable models, exploring novel multimodal fusion techniques, enhancing explainability of DL models, and addressing the ethical implications of automated content moderation. The continuous evolution of online content and dissemination methods necessitates ongoing innovation in detection strategies.