Phisher - A Multimodal Approach for Phishing Detection

Phishing is an illegal method used to trick people into revealing confidential information, such as login details, credit card numbers, and Social Security numbers.The majority of these phishing activities are carried out by duplicating the appearance of authentic websites or emails and exploiting people's trust, rather than technical vulnerabilities.Numerous awareness campaigns and technical countermeasures are designed to alert individuals to the dangers of phishing.Still, it remains one of the most effective methods of cyber assault due to its malleability and continually evolving complexity.Many single-modal models are effective to a certain degree, but cannot identify advanced phishing techniques that incorporate dynamic web content, obfuscated scripts, and sophisticated visual mimicry.We introduced a novel multimodal approach called Phisher.Our multimodal models utilize the BERT Multimodal Large Language Model (MLLM) for combined lexical analysis, ResNet50 for image processing, and semantic characteristics for URL extraction, thereby enhancing phishing classification.By combining these signals, we can achieve better accuracy, precision, and F1 score, which facilitates more effective detection of phishing sites.To test our multimodal model, we utilized the TR-OP real-life dataset, which contains 10,000 labeled phishing and legitimate websites, including HTML content, URLs, and website snapshots.The results show a significant improvement in accuracy and precision compared to other models.Aside from the technical benefits, this research also demonstrates how Multimodal learning can create more resilient defenses against evolving cybercrimes and phishing and offer practical applications for enterprises and security providers to build a safer digital ecosystem.