Table 4 Comparison of AFS with other methods on forgetting QF and model performance with the PathMNIST dataset

From: A unified method to revoke the private data of patients in intelligent healthcare with audit to forget

Methods

QF100

QF1000

Accuracy

F1-score

Independent teacher

1

1

0.8538

0.9885

Independent student

1

1

0.8446

0.9836

Independent teacher (k = 0.75)

1.08e−03

3.11e−30

0.8214

0.9796

Independent student (k = 0.75)

2.35e−02

4.08e−15

0.8396

0.9555

AFS w/o Audit (k = 0.75)

1.08e−03

1.67e−22

0.8682

0.9777

AFS (k = 0.75)

2.25e−05

2.05e−41

0.8560

0.9605

Independent teacher (k = 0.5)

3.74e−03

2.91e−23

0.7100

0.8314

Independent student (k = 0.5)

6.91e−03

2.99e−21

0.7934

0.9533

AFS w/o Audit (k = 0.5)

3.74e−03

4.93e−18

0.8494

0.9697

AFS (k = 0.5)

2.87e−06

4.75e−35

0.8242

0.9575

Independent teacher (k = 0.25)

3.74e−03

2.52e−26

0.7026

0.8282

Independent student (k = 0.25)

1.58e−07

9.05e−57

0.7582

0.9287

AFS w/o Audit (k = 0.25)

3.32e−07

2.05e−41

0.7842

0.9406

AFS (k = 0.25)

3.32e−07

1.84e−56

0.7810

0.9385

SISA (10 shards, QF excluded, k = 0.99 for QF100 and 0.9 for QF1000

2.15e−03

6.73e−29

0.8501

0.9840

CF-k (QF excluded)

1

1.69e−02

0.8506

0.9839

EU-k (QF excluded)

2.35e−02

7.69e−16

0.8328

0.9596

  1. QF100 is the small query dataset containing 100 samples and QF1000 is the large query dataset containing 1000 samples. We present the p values of auditing models trained with different methods on QF100 and QF1000 and the model performance including the accuracy and F1-score. p values were calculated using two-tailed Student’s t test.
Back to article page