Abstract
Voice assistants (VAs) rely on wake-up words to initiate interaction, but they are frequently triggered by phonetically similar "fuzzy words." This paper introduces FakeWake, a systematic framework to automatically generate these triggers, understand their phonetic causes, and mitigate the resulting privacy risks. By evolving fuzzy words using a genetic algorithm and identifying "decisive factors" through interpretable models, the researchers generated nearly 1,000 new triggers. Furthermore, they demonstrate that adversarial retraining can reduce accidental activations by over 90% while simultaneously improving the detection of legitimate wake-up words.
The FakeWake Problem
Voice assistants like Alexa, Siri, and Google Assistant use a "wake-up word" mechanism to balance power efficiency with user privacy. They listen locally for a specific keyword before recording and sending audio to the cloud. However, these systems are prone to the FakeWake phenomenon: accidental triggers caused by words that sound similar to the wake-up word but are linguistically distinct (e.g., "unacceptable" triggering "Alexa").
These false positives aren't just annoying; they are a major privacy risk. Studies show that 50% of users experience accidental wake-ups at least once a week, often resulting in private conversations being recorded and uploaded. Previous research relied on manual, exhaustive searches through media to find these words. This paper seeks to automate the discovery of fuzzy words, understand the underlying phonetic causes, and build more resilient detectors.
The FakeWake Framework
The authors approached the problem through a systematic three-phase framework designed to generate, understand, and defend against these accidental triggers.
- Automated Generation: Instead of searching through existing audio, they built a fuzzy word generator using a Genetic Algorithm (GA). Words are encoded as vectors (Pinyin for Chinese, phonemes for English). The GA "evolves" words through mutation and crossover, optimizing for a high wake-up rate and high dissimilarity from the actual wake-up word.
- Interpretable Understanding: Since commercial VAs are "black boxes," the authors built an interpretable tree-based proxy model to simulate VA behavior. They used SHAP (SHapley Additive exPlanations) values to pinpoint "decisive factors"—specific phonetic snippets that contribute most to a false acceptance.
- Mitigation: They proposed adversarial retraining, where the wake-up word detector is retrained using the newly discovered fuzzy words as negative samples. This helps the model more precisely define the decision boundary between legitimate triggers and similar-sounding noise.
Key Insights and Results
Phonetic Anchors and the "Fewer is Worse" Rule
The study revealed that VAs don't listen to the entire wake-up word with equal weight. They rely on "decisive factors"—short phonetic snippets (like the "ks" sound in "Alexa"). Wake-up words that rely on fewer decisive factors are much easier to spoof. For example, AliGenie relies heavily on the "ian" sound, leading to a significantly higher number of fuzzy words compared to more phonetically complex triggers.
Efficiency and Mitigation Success
- Speed: The framework generated 965 fuzzy words in just 4 hours—a task that previously took researchers weeks of manual searching.
- Defense Efficacy: Retraining models with fuzzy words was highly effective. For the "Alexa" model, the fuzzy rate dropped from 22.89% to 0.25%.
- Performance Boost: Surprisingly, the strengthened models didn't just ignore fuzzy words better; they actually became more accurate at identifying the real wake-up words. Training on samples near the decision boundary helped the model refine its understanding of the target keyword.
Environmental Stability
Fuzzy words with high wake-up rates tend to remain effective even when environmental factors like volume, speed, or background noise change. Human trials with 33 volunteers confirmed that the generated fuzzy words sounded significantly different from the real wake-up words, highlighting a critical gap between human and machine perception.
Conclusion
FakeWake provides a robust toolset for identifying and fixing vulnerabilities in voice assistant wake-up word detection. By bridging the gap between human and machine perception, this framework offers a path toward more secure and private smart home environments.
Summary updated