Explanation APIs are often treated as increasing model-extraction risk, but an end-to-end attack comparison cannot show whether an explanation reveals model-specific information or merely supplies a convenient next-query coordinate. We study this distinction for explanation-guided extraction of tabular classifiers. We separate an explanation into an attribution channel, which ranks influential features, and a coordinate channel, which supplies candidate feature values for perturbing those features. We then test whether the useful bin edges are privileged to the service or reproducible from the adversary's own data. Across six datasets and five target families, using attribution rankings to choose which features to perturb provides no systematic extraction advantage. Lime's discretization bin edges can improve fidelity by up to 41 percentage points on continuous features, but those edges are quantiles of the explainer's background data rather than properties of the target model. In controlled comparisons, an attacker-side grid reproduces the useful bin-edge gains without querying an explanation API. We use these findings to derive Cutpurse, which combines a self-computed grid with diverse query generation and boundary search. Against a service that withholds bin edges, Cutpurse improves fidelity by 16.3 percentage points on continuous-feature targets. It also exceeds the undefended published attack by 7.0 percentage points. Thus, withholding default Lime bin edges alone does not prevent effective extraction under the in-domain auxiliary-input, probability-output, and same-family surrogate assumptions studied here. Our results concern default Lime and Shap configurations for tabular classifiers. They do not establish the same conclusion for other explainers, target classes, or data-free attackers.
Figures & tables
Fig. 1: The attacker fits a grid on auxiliary data, then uses diverse generation, boundary search, and self-grid traversal to construct a surrogate dataset. It queries the prediction API but does not query the explanation API, the greyed-out explanation endpoint is shown for context only
n
feat.
cont.
cls.
Q
crop
1,700
7
7
17
1000
adult
32,561
11
3
2
1000
breast
569
30
30
2
100
nursery
12,960
8
0
3
1000
mushroom
8,123
21
0
2
1000
pendigits
10,992
16
16
10
1000
TABLE I: Datasets. cont. counts continuous features and Q is the query budget. Train, evaluation and auxiliary partitions follow the 75:15:10 split of the attack we analyse.
crop
pendigits
breast
adult
nursery
mushroom
LR
−5.5
−1.9
+1.2
−0.0
−10.2
−11.2
NB
−1.0
+0.6
+1.4
+2.5
−9.1
+3.8
KNN
+1.0
+1.0
−0.3
+1.0
−4.6
+3.6
DT
+1.3
+2.7
+2.3
+0.7
−1.1
+9.8
RF
+0.4
−1.5
+0.3
+0.7
+0.7
+6.9
TABLE II: RQ1: SHAP-ranked versus random feature selection. Mean paired fidelity difference (pp). Positive values mean SHAP-ranked selection achieves higher fidelity.
crop
pendigits
breast
adult
nursery
mushroom
LR
−3.8
+0.1
+6.0
−4.5
−0.2
−2.9
NB
+6.8
−1.4
+5.2
−1.8
−6.0
−15.4
KNN
+4.4
−0.6
+4.0
−0.3
+5.7
+6.1
DT
−2.2
−5.7
−2.7
−6.0
+1.5
+5.3
RF
−4.0
−6.9
−5.6
−3.7
+1.5
+7.7
TABLE III: RQ1: LIME-ranked versus random feature selection. Mean paired fidelity difference (pp). Bin edges are available to both configurations. Positive values mean LIME-ranked selection achieves higher fidelity.
crop
pendigits
breast
adult
nursery
mushroom
LR
+40.9
+18.6
−10.2
−1.8
−0.8
−8.7
NB
+24.9
+7.9
+11.2
+0.0
+5.0
−0.2
KNN
+15.2
+0.4
−0.2
−0.3
+1.0
+13.9
DT
+38.0
+10.1
+4.8
+2.6
+0.6
+2.1
RF
+22.7
+0.7
+1.1
+0.9
+0.5
−0.8
TABLE IV: RQ2: Effect of LIME bin edges. Mean paired fidelity difference (pp). Positive values mean bin-edge traversal achieves higher fidelity.
dataset
model
tgt
aux
target adv.
crop
LR
+40.9
+41.9
−0.9
crop
DT
+38.0
+36.8
+1.1
crop
NB
+24.9
+24.1
+0.9
crop
RF
+22.7
+22.4
+0.3
pendigits
LR
+18.6
+20.1
−1.5
crop
KNN
+15.2
+17.7
−2.5
TABLE V: RQ3: Target and auxiliary LIME grids. Fidelity gain over no-bin-edge traversal (pp). Target advantage =\textsctgt−\textscaux , so negative values favour the auxiliary grid.
dataset
model
target
n=1
n=2
crop
LR
+40.9
+40.7
+43.9
crop
DT
+38.0
+38.2
+38.2
pendigits
LR
+18.6
+12.9
+15.4
pendigits
DT
+10.1
+9.3
+9.3
TABLE VI: RQ3: Quartile grids fitted on a small initial pool. Fidelity gain over no-bin-edge traversal (pp). Target uses target training features. n is the number of initial inputs per class used to fit the auxiliary grid. Full results are in Appendix D .
TABLE VIII: RQ3: Entropy-grid results with two initial inputs per class. Fidelity gain over no-bin-edge traversal (pp).
fidelity
contrast
dataset
model
blind
defended
autolycus
self-grid
Cutpurse
sg − aut
Cutpurse − aut
Cutpurse − def
crop
LR
38.9
39.5
80.4
82.3
91.6
+1.8
+11.1
+52.1
crop
NB
57.5
56.0
80.9
72.7
75.4
−8.2
−5.5
+19.5
crop
KNN
53.6
51.4
66.6
61.4
63.1
−5.3
−3.5
+11.6
crop
DT
40.0
39.7
77.6
77.9
89.9
+0.3
+12.3
+50.3
crop
RF
64.3
59.5
82.2
86.1
91.3
+3.9
+9.2
+31.9
TABLE IX: Evaluation of Cutpurse . Mean fidelity (%) and pairwise contrasts (pp). sg − aut compares the auxiliary grid and random feature selection with the target explanation under the base traversal. self-grid and Cutpurse make no explanation calls.
Fig. 2: Fidelity vs. query budget for Autolycus (grey, dashed), Autolycus with bin edges withheld (red, dotted) and Cutpurse (blue, solid), per dataset (rows) and target (columns), for the four datasets with continuous features. Band is the standard error over ten independent target instantiations. Same splits, seed sets and attack RNG as Table IX , whose cells are the rightmost point of each curve.
Appendix figures & tables9 assets
Supplementary material from the paper’s appendix.
Appendix
Fig. 3: Contribution of each component to the LIME fidelity gain over Autolycus (percentage points, mean over datasets, top budget). Boundary search dominates for RF/LR. Diverse generation dominates for NB.
Fig. 4: Per dataset: contribution of boundary search (Phase 2, orange) and diverse generation (Phase 1, green) to the LIME fidelity gain over Autolycus (percentage points, top budget, mean over S=10 seed sets), one panel per dataset, grouped by target model. Shared y -axis. The adult /NB diverse bar ( +34 pp) sets the scale.
Fig. 5: Fidelity vs. query budget, our phases reading the target’s explanation (solid, blue) vs. Autolycus- Lime (dashed, grey), per target (rows) and dataset (columns). Band is the standard error over S=10 seed sets. n=1 , k=3 , diverse budget fixed at cap =15 . Shared y -axis ( 0.4 – 1.0 ).
contrast
mean
95% interval
Cutpurse − Autolycus, feature-scaled
+5.48
[+1.55,+10.68]
Cutpurse − Autolycus, unit
+3.96
[+2.04,+5.86]
Cutpurse − stronger Autolycus rule
+3.29
[+0.94,+5.81]
Autolycus: feature-scaled − unit
+1.45
[−3.77,+4.91]
Appendix
TABLE X: Step-size sensitivity on continuous-feature targets. Mean fidelity difference (pp) over 20 dataset–model cells.
aux , fitted on n per class
dataset
model
target
n=1
n=2
n=3
n=5
full aux
crop
LR
+40.9
+40.7
+43.9
+40.9
+42.8
+41.9
crop
DT
+38.0
+38.2
+38.2
+38.6
+37.3
+36.8
crop
NB
+24.9
+23.4
+23.4
+24.5
+24.1
+24.1
crop
RF
+22.7
+20.5
+20.7
+23.4
+22.1
+22.4
pendigits
LR
+18.6
+12.9
+15.4
+16.5
+16.6
+20.1
Appendix
TABLE XI: Full quartile-grid sensitivity sweep. Fidelity gain over no-bin-edge traversal (pp) with the auxiliary grid fitted on n inputs per class. The target column is the target-grid gain. Ten paired splits.
dataset
model
moderate
severe
crop
LR
−2.5
+3.5
crop
NB
+3.3
+2.8
crop
KNN
+1.1
+3.5
crop
DT
−0.9
+6.2
crop
RF
+4.1
+4.1
pendigits
LR
+2.8
+6.5
Appendix
TABLE XII: Class-support sensitivity by target. Target-grid advantage (pp); positive values favour the target grid.
effect
moderate
severe
restricted grid only
+0.1
−1.3
restricted seeds only
−9.3
−20.2
restricted grid and seeds
−11.0
−26.7
interaction
−1.8
−5.3
own-grid bin-edge gain
+22.3
+20.4
Appendix
TABLE XIII: Class-support decomposition. Mean fidelity difference (pp) over eight cells.
Fig. 6: Defense sweep on the two purely categorical datasets, the companion to Figure 2 . Withholding the bin edges costs Autolycus almost nothing here, since integer-encoded categories give the discretizer no meaningful cut points, and Cutpurse has no edge to snap to either.
Fig. 7: Shap companion: surrogate fidelity vs. query budget. Our phases reading the target’s Shap output (solid, blue) vs. Autolycus- Shap (dashed, grey), per target (rows) and dataset (columns). Shaded band is the standard error over the S=10 seed sets. n=5 , k=3 . All panels share a fixed y -axis ( 0.4 – 1.0 ) for cross-panel comparability.
Department of Statistics and Operations Research, University of North Carolina at Chapel Hill · Department of Statistics, University of California, Irvine