Abstract
With the growing development of the Internet, phishing sites have caused increasingly severe harm, often resulting in irretrievable economic losses for users. This paper proposes a hybrid detection model that combines a Perceived Hash Algorithm (PHA) with web page "noise" analysis. The model first utilizes PHA to compare the visual similarity of web page screenshots. For cases where visual similarity is inconclusive, the system extracts structural "noise" (irrelevant HTML elements) and employs N-Gram modeling to generate feature vectors. These features are then compared against a corpus of known sites using cosine similarity. Experimental results demonstrate that this hybrid approach is more effective and efficient than traditional multi-feature machine learning models, achieving higher precision and recall in detecting sophisticated phishing clones.
Introduction
Phishing remains one of the most pervasive forms of cybercrime because it targets the human element of security. By creating visually identical replicas of trusted websites, attackers deceive users into revealing sensitive information such as login credentials and financial data. The core challenge in detection is the "relevance" problem: phishing pages are designed to be indistinguishable from their legitimate counterparts.
Existing detection methods generally fall into four categories: URL blacklists, URL-based machine learning, content-based detection, and IP address tracking. However, these methods face significant limitations. Blacklists are inherently reactive and fail to stop "zero-day" phishing sites. Content-based detection, while more robust, is often computationally expensive and can be bypassed by minor textual variations or the use of images to represent text. This paper introduces a tiered model that leverages both visual "fingerprinting" and structural analysis to provide a more resilient defense.
Proposed Methodology
The proposed model operates as a tiered pipeline to balance computational efficiency with detection accuracy.
1. Visual Similarity via Perceived Hash Algorithm (PHA)
The first stage of the model involves a visual check. If a URL is unknown, the system captures a screenshot of the website and generates a 64-bit hash using PHA. Unlike cryptographic hashes, PHA is designed to be "perceptually" similar; small changes in the image result in small changes in the hash.
The process involves shrinking the screenshot to an 8x8 grid, converting it to grayscale, and binarizing the pixels based on the average brightness of the entire image. The similarity between the test site and protected sites is measured using Hamming distance ($d$):
- $d \le 5$: The site is classified as a definite phish.
- $d > 10$: The site is considered safe.
- $5 < d \le 10$: The site enters an "uncertainty zone," triggering the second stage of analysis.
2. Structural Analysis via Web Page Noise
For sites in the uncertainty zone, the model analyzes "web noise." The authors define noise as HTML elements that do not contribute to the main subject matter, such as navigation bars, footers, copyright notices, and advertising blocks. While most content-based filters attempt to remove this noise, this model treats it as a stable fingerprint. Because phishers often copy the entire source code of a target page to maintain visual fidelity, they inadvertently copy this structural noise.
The extracted noise is processed using a Bi-Gram model to create a feature vector. The system then calculates the Cosine Similarity between the test site's noise vector and the noise vectors of protected sites in the whitelist.
Experimental Results and Discussion
The researchers evaluated the model using data from PhishTank, focusing on six high-value targets: Adobe, AOL, Facebook, Google, Apple, and PayPal. The official sites for these brands served as the whitelist.
The experiment compared the performance of the visual-only PHA stage against the full hybrid model. The results show a marked improvement when structural noise is included.
| Metric | PHA Detection (Visual Only) | Web Noise Detection (Hybrid) |
|---|---|---|
| Precision | 71.43% | 88.26% |
| Recall | 70.97% | 85.26% |
| F-Score | 71.20% | 86.73% |
The hybrid model achieved a precision of 88.26%, compared to 71.43% for the PHA-only method. Furthermore, the researchers found that phishing sites typically exhibit a noise similarity score concentrated between 0.69 and 0.75, providing a reliable threshold for automated classification.
Conclusion
This paper demonstrates that "web noise," often discarded in other detection models, is a highly effective signal for identifying phishing clones. By combining visual PHA checks with structural N-Gram analysis, the proposed model achieves high detection accuracy while maintaining computational efficiency through its tiered architecture. Future research will focus on adapting the noise extraction process for modern, JavaScript-heavy web applications.
Summary updated