Abstract
Almost without any doubt, the Internet and Web in particular, have brought about radical changes in information retrieval with unparallel benefits to organizations and individuals to gain access to vast amount of articles and documents. The focus on restricting inappropriate materials at their source is not well suited to the nature and open architecture of Internet, where information source maybe in a different legal jurisdiction than the recipient. This paper proposes a reactive approach based on SOM (Self-Organising Maps) neural network to deal with dynamic blocking and filtering of Internet contents. We describe the design and implementation of a web policing proxy (WebPolice) based on a Kohonen's neural network model that attempts to classify the web contents dynamically using competitive learning. The parameter setting of the network has been experimented to obtain the optimal classification rate and performance for the model.
Introduction
In the early 2000s, the rapid expansion of the World Wide Web introduced significant challenges regarding content regulation, particularly for protecting minors from inappropriate material. Traditional methods like the Platform for Internet Content Selection (PICS) relied on voluntary self-labeling by content providers. However, this model proved ineffective as providers of restricted content lacked incentives to self-censor. This research introduces WebPolice, a reactive, recipient-centered filtering system that intercepts and classifies web content in real-time without requiring cooperation from content sources.
System Architecture and Methodology
The WebPolice system operates as a proxy server situated between the user's web browser and the internet. When a user requests a URL, the proxy fetches the page, extracts relevant keywords, and utilizes a neural network classifier to determine if the content is "Valid" or "Invalid."
The core classification engine is based on a Learning Vector Quantization (LVQ) network, a supervised variant of Kohonen’s Self-Organizing Map. The system uses a 207-dimensional input vector derived from a curated dictionary of keywords. Through competitive learning, the network adjusts "codebook vectors" (prototypes) for each category. If a winning neuron correctly identifies a sample, it moves closer to the input vector; otherwise, it is repelled.
Experimental Results and Key Insights
The effectiveness of the WebPolice system was evaluated using a custom dataset of HTML samples. A key finding was the impact of network complexity on classification accuracy. The researchers determined that the number of neurons (codebook vectors) per category significantly influenced performance.
As shown in Figure 6, the system's performance was tested across various configurations. The authors achieved a peak classification rate exceeding 64%. The optimal configuration was found to be $c=3$ neurons per category, suggesting that content categories are composed of distinct sub-clusters that a single prototype cannot represent. Furthermore, the study highlighted the importance of the learning rate; while a high initial rate ($ζ = 0.5$) was used in some experiments to observe convergence trends, a lower, more stable rate ($ζ = 0.05$) was found to facilitate better generalization.
Conclusion and Future Work
WebPolice demonstrates the feasibility of real-time, proxy-based content filtering using neural networks. While the 64% accuracy rate was a significant baseline for its time, several limitations remain. The current model is restricted to text-based analysis and is dependent on a static keyword dictionary. Future research will focus on incorporating image analysis, expanding the keyword store, and exploring more complex neural topologies to enhance classification precision in the evolving web landscape.
Summary updated