Radio-Frequency-Fingerprinting: Few-Shot-Specific-Emitter-Identification-via-Deep-Metric-Ensemble-Learning
Requirements: keras=2.1.4, tf=1.14.0
Paper: http://arxiv.org/abs/2207.06592 or Y. Wang, G. Gui, Y. Lin, H. -C. Wu, C. Yuen and F. Adachi, "Few-Shot Specific Emitter Identification via Deep Metric Ensemble Learning," in IEEE Internet of Things Journal, vol. 9, no. 24, pp. 24980-24994, 15 Dec.15, 2022, doi: 10.1109/JIOT.2022.3194967.
Change ADS-B 6000-> ADS-B 4800 (Remove the ICAO code) model weight, dataset and results are updated
A brief introduction to this code: (change 6000 to 4800)
- STC-CVCNN_Train: train feature embedding on auxiliary dataset of 90 classes, and visualization based on test dataset of 10 classes
- STC-CVCNN_Test: train LR classifer with few-shot training dataset (1-5-10-15-20 shots), and test it on test dataset. Here, this code executes 100 times, because different few-shot training datasets have different performance.
- STC-CVCNN_SC: feature visualization & get silhouette coefficient
New result （100 Monte Carlo simulations）
Different feature embedding with LR classifier
| C-K | FS CVCNN | Softmax | Siamese | Triplet | SR2CNN | *STC | ST | SC |
|---|---|---|---|---|---|---|---|---|
| 10-1 | 10.00% | 41.30% | 50.60% | 75.98% | 85.04% | 87.66% | 73.35% | 85.37% |
| 10-5 | 47.80% | 75.26% | 77.51% | 90.18% | 93.01% | 93.99% | 88.66% | 93.05% |
| 10-10 | 69.40% | 87.48% | 83.34% | 93.00% | 94.32% | 95.28% | 92.31% | 94.04% |
| 10-15 | 77.20% | 91.03% | 89.47% | 93.78% | 94.81% | 95.88% | 93.29% | 94.45% |
| 10-20 | 84.30% | 93.56% | 91.47% | 94.18% | 94.82% | 96.15% | 94.34% | 94.99% |
| 20-1 | 5.00% | 35.59% | 38.02% | 57.02% | 64.52% | 66.11% | 51.33% | 61.20% |
| 20-5 | 38.80% | 61.73% | 57.00% | 71.27% | 76.05% | 80.01% | 71.37% | 75.50% |
| 20-10 | 53.60% | 72.18% | 66.73% | 74.96% | 80.94% | 84.42% | 77.46% | 79.05% |
| 20-15 | 63.25% | 76.99% | 69.78% | 77.93% | 82.94% | 86.53% | 81.06% | 81.22% |
| 20-20 | 72.50% | 81.18% | 72.38% | 79.29% | 84.63% | 87.74% | 83.02% | 82.59% |
| 30-1 | 3.33% | 27.68% | 28.81% | 46.18% | 51.28% | 55.94% | 41.81% | 52.04% |
| 30-5 | 26.90% | 53.70% | 47.77% | 62.02% | 68.91% | 72.46% | 61.40% | 66.20% |
| 30-10 | 47.40% | 64.22% | 58.30% | 67.54% | 73.33% | 77.60% | 68.96% | 71.22% |
| 30-15 | 54.57% | 69.99% | 62.79% | 70.12% | 75.58% | 80.14% | 73.68% | 73.89% |
| 30-20 | 63.30% | 74.04% | 65.70% | 72.62% | 77.77% | 81.37% | 76.20% | 75.52% |
STC-based feature embedding with diffferent classifiers
| C-K | LR | LR-3models | LR-5models | LR-7models | KNN | RF | SVM |
|---|---|---|---|---|---|---|---|
| 10-1 | 87.66% | 89.07% | 90.12% | 89.84% | 21.12% | 78.21% | 86.75% |
| 10-5 | 93.99% | 95.22% | 95.62% | 95.54% | 91.86% | 93.81% | 93.48% |
| 10-10 | 95.28% | 96.35% | 96.53% | 96.48% | 93.92% | 94.70% | 94.33% |
| 10-15 | 95.88% | 96.97% | 97.05% | 97.06% | 94.73% | 95.19% | 94.88% |
| 10-20 | 96.15% | 97.23% | 97.40% | 97.40% | 95.25% | 95.37% | 95.22% |
| 20-1 | 66.11% | 69.33% | 69.97% | 71.17% | 22.76% | 58.72% | 65.24% |
| 20-5 | 80.01% | 82.77% | 83.56% | 83.96% | 73.51% | 78.29% | 77.26% |
| 20-10 | 84.42% | 87.61% | 87.83% | 88.19% | 80.67% | 83.33% | 81.65% |
| 20-15 | 86.53% | 89.95% | 90.22% | 90.52% | 83.58% | 85.41% | 84.45% |
| 20-20 | 87.74% | 91.41% | 91.34% | 91.63% | 85.49% | 87.13% | 86.04% |
| 30-1 | 55.94% | 60.40% | 60.89% | 61.69% | 21.74% | 48.27% | 54.74% |
| 30-5 | 72.46% | 77.12% | 77.87% | 78.35% | 64.81% | 70.42% | 68.95% |
| 30-10 | 77.60% | 82.28% | 82.93% | 83.63% | 72.53% | 76.01% | 74.52% |
| 30-15 | 80.14% | 84.85% | 85.40% | 85.88% | 75.36% | 78.57% | 77.51% |
| 30-20 | 81.37% | 86.36% | 86.79% | 87.48% | 77.62% | 80.19% | 79.26% |
STC-based feature embedding with different classifiers (PyTorch)
- The values in parentheses represent the 95% confidence interval.
| C-K | LR | KNN | SVM |
|---|---|---|---|
| 10-1 | 46.94 (±1.46) | 15.80 (±0.76) | 42.37 (±1.40) |
| 10-5 | 75.66 (±0.95) | 53.80 (±0.98) | 66.67 (±1.18) |
| 10-10 | 81.89 (±0.77) | 64.22 (±0.93) | 74.22 (±1.57) |
| 10-15 | 85.17 (±0.37) | 69.35 (±0.56) | 79.48 (±0.59) |
| 10-20 | 86.22 (±0.76) | 72.05 (±0.72) | 80.72 (±1.18) |
| 20-1 | 37.64 (±0.81) | 12.81 (±0.68) | 34.20 (±0.81) |
| 20-5 | 60.51 (±1.13) | 44.72 (±0.87) | 52.14 (±1.31) |
| 20-10 | 68.65 (±0.57) | 54.01 (±0.55) | 61.45 (±0.77) |
| 20-15 | 70.68 (±1.23) | 57.43 (±1.01) | 64.28 (±1.68) |
| 20-20 | 73.60 (±1.02) | 60.62 (±0.81) | 67.82 (±1.39) |
| 30-1 | 30.25 (±0.77) | 11.63 (±0.55) | 27.79 (±0.71) |
| 30-5 | 50.86 (±0.80) | 37.94 (±0.61) | 42.60 (±0.93) |
| 30-10 | 58.22 (±0.66) | 45.75 (±0.54) | 50.28 (±0.85) |
| 30-15 | 62.53 (±0.58) | 50.17 (±0.42) | 55.30 (±0.65) |
| 30-20 | 65.15 (±0.58) | 52.91 (±0.47) | 58.88 (±0.67) |
The influence of different sets of few-shot training samples (left: 10-way-shot with STC CVCNN and LR; right: 10-way-1-shot with STC CVCNN)
Model weight and Dataset
Link: https://pan.baidu.com/s/13qW5mnfgUHBvWRid2tY2MA Passwd：eogv
or