The package was missing every script behind the KM (lrn.) curves and
both learned table rows: exp_learned's driver, the merge that folds the
learned rows into sec_compare.csv and refresh_summary.csv, and the
report that reads the learned numbers back. It was also missing
sec_keylen_perm.csv, so Fig. 3 could not be regenerated at all, and the
two diagnostics that answer why a fixed key beats a learned one here and
where a learned mask would win instead.
The learned artifacts themselves are regenerated. They were trained on
the cross-entropy alone, which drifts to disjoint sparse supports: 99
percent of each key's energy on about six of the 64 entries, so a digit
is decided over a sixth of its period and the key set is a choice of
support rather than a dense direction in R^L. They are now the
regularized keys of Section V-C, and check_consistency asserts which of
the two families the figures draw.
Verified from inside this repository: replot_security.py rebuilds all
seven result figures, make_tables.py reproduces both result tables, and
check_consistency.py passes every check that does not need the
manuscript.
The README now lists what ships. Its run list, layout and figure map had
none of the learned pipeline, named two tables the manuscript renders as
prose, and gave Fig. 3 no data file for its permutation curve.
check_family_enum.py now runs both attacks against both families. The
outsider ranks the L-1 Walsh rows; the insider, holding m_v, ranks the
L-1 products m_v .* m_r, which works because Walsh rows are closed
under the elementwise product and the per-block sign cancels in
m_u .* m_v. Both need a list to rank, and only the structured family
supplies one: the structured family falls at 0.905 from one frame at
10 dB and 0.990 from four, the refresh takes the outsider to 0.000 and
leaves the insider at 0.980, and the learned family gives 0.000
throughout.
exp_full.stage_N sweeps the learned family over the same SNR grid at
the same frame count as stage_A, so Fig. 2 can carry both families and
a reader can see what the key space costs at every SNR rather than at
one point.
check_consistency.py gains four assertions for the key-space
measurements and two for the learned sweep, 82 in all.
README: the assertion count was two rounds stale, and the figure map
omitted family_enum, cov_attack and maskdegen, whose CSVs back quoted
manuscript numbers.
check_family_enum.py measures the attack the manuscript now states in
Section III-A: the winning correlation is an index-free verifier, so
ranking the 63 non-constant Walsh rows by mean winning correlation
recovers the user set from one frame in 0.905 of 200 trials at 10 dB
and from four frames in 0.990, using nothing outside the stated threat
model. Under the invariance refresh it recovers it in none, because the
entry permutation relabels the codebook the adversary must align
against.
V8 and V9 read the trained codebook through main_model(), which
retrains on every call, and a codebook trained on CUDA is not the one
trained on CPU. The shipped verify_math.csv therefore read PASS here
and FAIL for anyone running this package without a GPU. model_main.pt
is 7 KB and fixes the codebook, which is what both checks are about;
delete it to retrain. V1-V11 now pass on both.
New checks: V10, the format-matched OMA reference Section VI-B quotes,
and V11, the closed-form against Monte Carlo comparison the manuscript
claimed and never stored. V3a's bias-linearity result was computed and
printed but never written to the CSV, so the one linearity claim the
paper quotes was the one this package could not show.
check_consistency.py gains 21 assertions, covering five data files that
no assertion read (users, csi, semantic, cov_attack, sec_jam) and the
trend claims it structurally could not see, since it compared values
and not shapes.
README: the figure map named stages that do not write the artifacts
they list, so following it did not reproduce Figs. 4 and 6; the
reproduction block was five scripts short; and the refresh numbers were
from a superseded run (nearly three, 15.0 to 64.8 bits) against the
manuscript's 2.3 and 23.8 to 364.6.
KiHoLee