aprender-core 0.30.0

Next-generation machine learning library in pure Rust
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230

pub(crate) use super::*;

#[test]
fn test_fgsm_perturb() {
    let attack = FgsmAttack::new(0.1);
    let input = vec![1.0, 2.0, 3.0];
    let gradient = vec![0.5, -0.5, 0.0];

    let perturbed = attack.perturb(&input, &gradient);

    assert_eq!(perturbed.len(), 3);
    assert!((perturbed[0] - 1.1).abs() < 1e-6); // positive grad -> +epsilon
    assert!((perturbed[1] - 1.9).abs() < 1e-6); // negative grad -> -epsilon
    assert!((perturbed[2] - 3.1).abs() < 1e-6); // zero grad -> +epsilon (sign(0) = 1)
}

#[test]
fn test_fgsm_batch() {
    let attack = FgsmAttack::new(0.1);
    let inputs = vec![vec![1.0, 2.0], vec![3.0, 4.0]];
    let gradients = vec![vec![1.0, -1.0], vec![-1.0, 1.0]];

    let perturbed = attack.attack_batch(&inputs, &gradients);

    assert_eq!(perturbed.len(), 2);
    assert_eq!(perturbed[0].len(), 2);
}

#[test]
fn test_pgd_step() {
    let attack = PgdAttack::new(10, 0.01, 0.1);
    let original = vec![1.0, 2.0];
    let current = vec![1.05, 1.95];
    let gradient = vec![1.0, -1.0];

    let next = attack.step(&current, &original, &gradient);

    // Should move in direction of gradient sign
    assert!(next[0] > current[0]);
    assert!(next[1] < current[1]);

    // Should stay within epsilon of original
    assert!((next[0] - original[0]).abs() <= attack.epsilon + 1e-6);
    assert!((next[1] - original[1]).abs() <= attack.epsilon + 1e-6);
}

#[test]
fn test_pgd_attack() {
    let attack = PgdAttack::new(5, 0.01, 0.1);
    let input = vec![1.0, 2.0, 3.0];

    // Mock gradient function
    let result = attack.attack(&input, |_x| vec![1.0, 1.0, 1.0]);

    assert_eq!(result.len(), 3);
    // After attack, values should be perturbed up to epsilon
    for (orig, perturbed) in input.iter().zip(result.iter()) {
        assert!((perturbed - orig).abs() <= attack.epsilon + 1e-6);
    }
}

#[test]
fn test_gaussian_noise() {
    let attack = GaussianNoiseAttack::new(0.1, 42);
    let input = vec![1.0, 2.0, 3.0];

    let perturbed = attack.perturb(&input);

    assert_eq!(perturbed.len(), 3);
    // Values should be different from original
    let any_different = input
        .iter()
        .zip(perturbed.iter())
        .any(|(a, b)| (a - b).abs() > 1e-10);
    assert!(any_different);
}

#[test]
fn test_gaussian_noise_reproducibility() {
    let attack1 = GaussianNoiseAttack::new(0.1, 42);
    let attack2 = GaussianNoiseAttack::new(0.1, 42);
    let input = vec![1.0, 2.0, 3.0];

    let perturbed1 = attack1.perturb(&input);
    let perturbed2 = attack2.perturb(&input);

    // Same seed should give same results
    for (a, b) in perturbed1.iter().zip(perturbed2.iter()) {
        assert!((a - b).abs() < 1e-6);
    }
}

#[test]
fn test_attack_result() {
    let result = AttackResult::new("FGSM", 0.95, 0.92, 0.05, Duration::from_millis(100));

    assert_eq!(result.attack_name, "FGSM");
    assert!((result.accuracy_drop - 0.03).abs() < 1e-6);
    assert!(result.is_robust); // 3% < 5%

    let result2 = AttackResult::new("PGD", 0.95, 0.85, 0.05, Duration::from_millis(100));
    assert!(!result2.is_robust); // 10% > 5%
}

#[test]
fn test_adversarial_config_default() {
    let config = AdversarialConfig::default();
    assert!((config.fgsm_epsilon - 0.1).abs() < 1e-6);
    assert_eq!(config.pgd_steps, 10);
    assert!((config.max_accuracy_drop - 0.05).abs() < 1e-6);
}

#[test]
fn test_run_robustness_tests() {
    let config = AdversarialConfig::default();
    let (score, issues) = run_robustness_tests(&config);

    assert_eq!(score.points_possible, 20);
    assert!(score.tests_passed + score.tests_failed > 0);
    // PGD typically fails with default config
    assert!(!issues.is_empty());
}

#[test]
fn test_run_robustness_tests_all_robust() {
    // Create config where all tests pass (large max_accuracy_drop)
    let config = AdversarialConfig {
        max_accuracy_drop: 0.20, // 20% allowed drop - all should pass
        ..Default::default()
    };
    let (score, issues) = run_robustness_tests(&config);

    // With large allowed drop, all tests should pass
    assert!(score.tests_passed >= 3);
    assert!(issues.is_empty());
}

#[test]
fn test_run_robustness_tests_none_robust() {
    // Create config where all tests fail (tiny max_accuracy_drop)
    let config = AdversarialConfig {
        max_accuracy_drop: 0.001, // 0.1% allowed - all should fail
        ..Default::default()
    };
    let (score, issues) = run_robustness_tests(&config);

    // With tiny allowed drop, all tests should fail
    assert!(score.tests_failed >= 3);
    // Should have issues for FGSM, PGD, and noise
    assert!(issues.len() >= 2); // At least FGSM and PGD issues
}

#[test]
fn test_attack_result_negative_drop_clamped() {
    // Test that negative accuracy drop is clamped to 0
    let result = AttackResult::new(
        "Test",
        0.90,
        0.95, // attacked is higher than original
        0.05,
        Duration::from_millis(100),
    );

    assert!((result.accuracy_drop - 0.0).abs() < 1e-6);
    assert!(result.is_robust); // 0% drop is within threshold
}

#[test]
fn test_fgsm_perturb_empty_input() {
    let attack = FgsmAttack::new(0.1);
    let input: Vec<f32> = vec![];
    let gradient: Vec<f32> = vec![];

    let perturbed = attack.perturb(&input, &gradient);
    assert!(perturbed.is_empty());
}

#[test]
fn test_pgd_attack_zero_steps() {
    let attack = PgdAttack::new(0, 0.01, 0.1);
    let input = vec![1.0, 2.0, 3.0];

    // With zero steps, output should equal input
    let result = attack.attack(&input, |_x| vec![1.0, 1.0, 1.0]);
    assert_eq!(result, input);
}

#[test]
fn test_gaussian_noise_different_seeds() {
    let attack1 = GaussianNoiseAttack::new(0.1, 42);
    let attack2 = GaussianNoiseAttack::new(0.1, 43); // Different seed
    let input = vec![1.0, 2.0, 3.0];

    let perturbed1 = attack1.perturb(&input);
    let perturbed2 = attack2.perturb(&input);

    // Different seeds should give different results
    let any_different = perturbed1
        .iter()
        .zip(perturbed2.iter())
        .any(|(a, b)| (a - b).abs() > 1e-10);
    assert!(any_different);
}

#[test]
fn test_gaussian_noise_empty_input() {
    let attack = GaussianNoiseAttack::new(0.1, 42);
    let input: Vec<f32> = vec![];

    let perturbed = attack.perturb(&input);
    assert!(perturbed.is_empty());
}

#[test]
fn test_adversarial_config_clone() {
    let config = AdversarialConfig::default();
    let cloned = config.clone();

    assert!((config.fgsm_epsilon - cloned.fgsm_epsilon).abs() < 1e-6);
    assert_eq!(config.pgd_steps, cloned.pgd_steps);
}

#[test]
fn test_attack_result_clone() {
    let result = AttackResult::new("Test", 0.95, 0.90, 0.05, Duration::from_millis(100));
    let cloned = result.clone();

    assert_eq!(result.attack_name, cloned.attack_name);
    assert!((result.accuracy_drop - cloned.accuracy_drop).abs() < 1e-6);
}