Abstract
Sound-squatting is a phishing attack that tricks users into malicious resources by exploiting similarities in the pronunciation of words. Proactive defense against sound-squatting candidates is complex, and existing solutions rely on manually curated lists of homophones. We here introduce Sound-skwatter, a multi-language AI-based system that generates sound-squatting candidates for proactive defense. Sound-skwatter relies on an innovative multi-modal combination of Transformers Networks and acoustic models to learn sound similarities. We show that Sound-skwatter can automatically list known homophones and thousands of high-quality candidates. In addition, it covers cross-language sound-squatting, i.e., when the reader and the listener speak different languages, supporting any combination of languages. We apply Sound-skwatter to network-centric phishing via squatted domain names. We find ~ 10% of the generated domains exist in the wild, the vast majority unknown to protection solutions. Next, we show attacks on the PyPI package manager, where ~ 17% of the popular packages have at least one existing candidate. We believe Sound-skwatter is a crucial asset to mitigate the sound-squatting phenomenon proactively on the Internet. To increase its impact, we publish an online demo and release our models and code as open source.
Introduction
While typosquatting (e.g., gogle.com) is a well-known threat, sound-squatting represents a more subtle and growing danger. This attack vector exploits phonetic similarities to deceive users, particularly in environments dominated by voice assistants, smart speakers, or cross-language interactions where a user might hear a brand name and phonetically misspell it when typing.
Traditional defenses against sound-squatting are largely reactive and manual, relying on static lists of homophones (e.g., "new" vs. "knew"). These lists fail to capture "quasi-homophones"—invented words that sound like legitimate ones—and do not account for cross-language phonetic interpretations. Sound-skwatter addresses these gaps by automating the generation of high-quality sound-alike candidates using a multi-modal AI approach.
Methodology
Sound-skwatter treats the generation of sound-squatting candidates as a sequence-to-sequence translation task. The system architecture is built upon two core pillars: phonetic representation and multi-modal training.
Phonetic Representation
The system uses the International Phoneme Alphabet (IPA) as a linguistic bridge. Target words are first converted into IPA sequences using the eSpeak NG engine, providing a standardized representation of sound regardless of the original language's orthography.
Multi-modal Training
The "secret sauce" of Sound-skwatter is its multi-modal training objective. As shown in the training architecture below, the model is not only trained to map IPA sequences back to text (graphemes) but is also tasked with reconstructing the Mel Spectrogram of the word.
By incorporating acoustic feedback as a supervisory signal, the model learns the subtle nuances of how different character combinations produce similar auditory results. During inference, the system employs a Beam Search algorithm to explore and rank the most probable alternative spellings for a given phonetic sequence.
Key Insights: The Power of Acoustic Feedback
A central finding of this research is that acoustic signals are essential for capturing sound similarities that purely text-based models miss. The inclusion of Mel Spectrogram reconstruction allows the model to identify phonetic overlaps in words like "silvester" and "silvestre" more effectively.
As illustrated in the CDF plot, the "Feedback" model (using acoustic signals) significantly outperforms the "No feedback" and "Random" baselines in identifying known homophones within the top-ranked candidates, demonstrating the value of the multi-modal approach.
Results and Real-World Impact
The effectiveness of Sound-skwatter was validated across several domains:
- Homophone Discovery: The system automatically generated 83.91% of known English homophones.
- Domain Squatting: When applied to the top 200 web domains, ~10% of the generated candidates were already registered in the wild. Crucially, most of these were not flagged by existing security services like Google Safe Browsing.
- Supply Chain Security: Analysis of the PyPI repository revealed that 16.72% of the top 5,000 packages had at least one sound-alike candidate already present, indicating a significant risk for developers.
- Cross-Language Defense: The model can predict how speakers of one language might phonetically spell words from another (e.g., an English speaker spelling the Italian comunità), enabling global brands to proactively secure their identity.
Conclusion
Sound-skwatter provides a powerful, automated tool for proactive defense against phonetic phishing. While currently focused on American English and limited to 20-character strings, its language-agnostic architecture paves the way for broader linguistic coverage. By open-sourcing the models and code, the authors aim to empower the security community to mitigate sound-squatting before it becomes a dominant attack vector.
Summary updated