pla-disc-results / README.md
ArchCoder's picture
Add DISC micro-Doppler data with CC BY 4.0 attribution
41b42d7 verified
|
Raw History Blame Contribute Delete
3.83 kB
metadata
license: cc-by-4.0
pretty_name: PLA on DISC - experiment results and sample index
task_categories:
  - time-series-forecasting
tags:
  - physical-layer-authentication
  - mmwave
  - isac
  - micro-doppler
  - iot
size_categories:
  - n<1K

PLA on DISC: experiment results and sample index

Companion data for the code in Deep-Learning-Based-Physical-Layer-Authentication-for-ISAC-Enabled-IoT-System.

The micro-Doppler spectrograms in data/ are taken unmodified from the DISC dataset (Pegoraro et al., IEEE DataPort, DOI 10.21227/2gm7-9z72, CC BY 4.0), and preprocessed_md.h5 is derived from them. Attribution and a list of changes are in DATA_LICENSE.md. On top of that data this repository holds everything we produced: the per-run result files, the training logs, and an index that says which DISC recording is which sample in our experiments.

Contents

data/micro_doppler_stft/  433 DISC spectrograms (unmodified)
data/preprocessed_dataset/preprocessed_md.h5   (432, 64, 64, 1) model input
metadata/samples.csv     one row per DISC micro-Doppler file (433 rows)
results/                 JSON output of every run (metrics, confusion matrices, losses)
logs/                    stdout of the five-split and four-seed runs behind Table 4

metadata/samples.csv

column meaning
sample_id DISC file name without .npy, e.g. WALKING_3_17
activity HANDS, RUNNING, WALKING or SITTING
subject subject id, 1 to 7
index recording number for that subject and activity
binary_label label used for authentication: HANDS, RUNNING -> 0; WALKING, SITTING -> 1
used_in_experiments 0 for SITTING_3_12, which is entirely NaN, otherwise 1

That leaves 432 usable samples: 133 with label 0 and 299 with label 1. By activity: walking 222, sitting 77, running 73, waving hands 60.

results/

ablation_results_R_sp<S>_sd<N>.json is one run with split seed S and training seed N. The five files with sd42 and sp0..sp4 are the five splits behind the mean ± sd in Table 4, and sp0 with seeds 7, 13, 42 and 99 shows the effect of the training seed. Each file holds, for every ablation row, accuracy, precision, recall, F1, FPR, the confusion matrix and the chosen k, plus the CDAE loss history. table3_md.json is the classifier comparison and TABLE4_final.json the aggregated table. Files without the R_ prefix come from earlier experiments and are kept for completeness.

Rebuilding the model input

python preprocess_md.py in the code repository rebuilds preprocessed_md.h5 from data/micro_doppler_stft/; the output is bit-identical to the file here. The raw .mat channel impulse responses (about 68 GB) are not included; get them from IEEE DataPort.

Reported results

Mean ± sd over five stratified splits, 200 epochs:

Method ACC F1 FPR
Raw micro-Doppler + KNN 87.82 ± 1.6 91.08 ± 1.2 20.71 ± 3.5
CDAE + KNN 88.51 ± 2.2 91.19 ± 1.9 10.71 ± 5.1
CDAE + SVM 92.18 ± 3.1 94.15 ± 2.5 10.71 ± 4.5
CDAE + Weighted KNN 93.56 ± 2.9 95.23 ± 2.1 9.29 ± 4.8

The test set has 28 negatives, so FPR changes in 3.57-point steps. GPU runs are not bit-reproducible; compare means rather than single runs.

Licence and citation

Everything here is CC BY 4.0. The DISC data is the work of its authors; please cite it:

Pegoraro, Saucedo, Lacruz, Rossi, Widmer. DISC: a Dataset for Integrated Sensing and Communication in mmWave Systems. IEEE DataPort, doi:10.21227/2gm7-9z72.