First-hand information for everyone
A new authentication framework, LABBA, combines a simple phone‑blowing sound with a lightweight large language model (LLM) to decide whether the user is genuine. The system extracts spectral and temporal cues from the blow‑acoustic signal and turns them into similarity scores that the LLM uses with structured prompts to make a context‑aware decision.
LABBA was tested with 50 participants under varied conditions. In the rule‑based mode the system reached 99.46% accuracy, while the score‑based mode achieved 99.38% accuracy, both higher than the 99.09% accuracy of the baseline that does not use the LLM.
The experiment also measured false rejection and false acceptance rates. Rule‑based LABBA had an 18.27 % FRR and 0.21 % FAR; its baseline had a 43.64 % FRR and 0.13 % FAR. Score‑based LABBA showed a 13.80 % FRR and 0.38 % FAR, again outperforming the baseline.
When the two LLM modes were combined, the minimum equal error rate (EER) dropped to 9.24 % for rule‑based fusion and 7.09 % for score‑based fusion, compared with 21.89 % for the baseline.
| Method |
|---|
| Gaussian |
|---|
| Accuracy (%) |
|---|
| FAR (%) |
|---|
| FRR (%) |
|---|
| EER (%) |
|---|
| Without LLM-Assisted | NA | 99.09 | 0.13 | 43.64 | 21.89 |
|---|
| With LLM-Assisted (Rule-Based) | Fast | 99.32 | 0.20 | 26.78 | 13.49 |
| Slow | 99.33 | 0.22 | 25.20 | 12.71 |
| Fused | 99.46 | 0.21 | 18.27 | 9.24 |
| With LLM-Assisted (Score-Based) | Fast | 99.25 | 0.58 | 14.22 | 7.40 |
| Slow | 99.22 | 0.53 | 14.4 | 7.48 |
| Fused | 99.38 | 0.38 | 13.8 | 7.09 |
The research prototype runs on a CPU and takes 8.237 s for rule‑based and 11.604 s for score‑based inference. These times are higher than the preceding feature‑extraction steps and would need optimisation for real‑time use.
| Method | Time (s) |
| Spectral Feature Extraction | 0.589 |
| Temporal (Fast-Smoothed) Feature Extraction | 0.486 |
| Temporal (Slow-Smoothed) Feature Extraction | 2.350 |
| Temporal (Fused-Smoothed) Feature Extraction | 2.832 |
| Feature Aggregation | 3.2×10−5 |
| Prompt Generation | 8.7×10−7 |
| LLM Evaluation (Rule-Based) | 8.237 |
| LLM Evaluation (Score-Based) | 11.604 |
The authors note that even with a lightweight LLM the system can still misclassify users, and that relying on structured prompts may lead to hallucinations. These limitations suggest further study is needed before deployment.
Overall, LABBA shows that a conversational AI module can improve the accuracy of a behavioral biometric that is simple, liveness‑safe and easy to use, though latency and decision fidelity remain practical hurdles.