optirs-core 0.3.2

OptiRS core optimization algorithms and utilities
Documentation 
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

//! Scoring, reputation and behaviour tracking

use crate::error::Result;
use scirs2_core::ndarray::Array1;
use scirs2_core::numeric::Float;
use std::cmp::Ordering;
use std::collections::HashMap;
use std::fmt::Debug;

use super::aggregator::ByzantineTolerantAggregator;
use super::helpers::{cosine_similarity, euclidean_distance, from_scalar, l2_norm};
use super::types::{
    AnomalyScore, BehaviorHistory, OutlierScore, ReputationScore, TrustLevel, VerificationScore,
};

impl<T: Float + Debug + Send + Sync + 'static + scirs2_core::ndarray::ScalarOperand>
    ByzantineTolerantAggregator<T>
{
    /// Combined Byzantine score in `[0, 1]`; every input is already bounded to
    /// `[0, 1]`, so the weights below are a genuine convex combination.
    pub(super) fn compute_byzantine_score(
        &self,
        anomaly_score: &AnomalyScore,
        outlier_score: &OutlierScore,
        verification_score: Option<&VerificationScore>,
    ) -> f64 {
        let mut combined_score = anomaly_score.score.clamp(0.0, 1.0) * 0.4;
        combined_score += outlier_score.score.clamp(0.0, 1.0) * 0.3;
        if let Some(verification) = verification_score {
            combined_score += (1.0 - verification.score.clamp(0.0, 1.0)) * 0.3;
        }
        combined_score
    }
    /// Confidence in the round's aggregate, in `[0, 1]`.
    ///
    /// The product of two measured quantities: the fraction of the submitting
    /// cohort that survived detection, and how tightly the survivors cluster around
    /// the aggregate (their mean distance to it, relative to the median update
    /// norm).
    pub(super) fn calculate_confidence_score(
        &self,
        honest_participants: &HashMap<String, Array1<T>>,
        total_submitted: usize,
        aggregate: &Array1<T>,
    ) -> Result<f64> {
        if total_submitted == 0 || honest_participants.is_empty() {
            return Ok(0.0);
        }
        let survivor_fraction =
            (honest_participants.len() as f64 / total_submitted as f64).min(1.0);
        let ordered = Self::ordered_cohort(honest_participants);
        let mut dispersion = 0.0f64;
        let mut norms: Vec<f64> = Vec::with_capacity(ordered.len());
        for (_, gradient) in &ordered {
            dispersion += from_scalar(euclidean_distance(gradient, aggregate)?)?;
            norms.push(from_scalar(l2_norm(gradient))?);
        }
        dispersion /= ordered.len() as f64;
        norms.sort_by(|a, b| a.partial_cmp(b).unwrap_or(Ordering::Equal));
        let scale = if norms.len().is_multiple_of(2) {
            let mid = norms.len() / 2;
            (norms[mid - 1] + norms[mid]) / 2.0
        } else {
            norms[norms.len() / 2]
        };
        let consensus = if scale > 0.0 {
            1.0 / (1.0 + dispersion / scale)
        } else if dispersion == 0.0 {
            1.0
        } else {
            0.0
        };
        Ok((survivor_fraction * consensus).clamp(0.0, 1.0))
    }
    /// Record this round's observed behaviour for every submitting participant.
    pub(super) fn record_behavior(
        &mut self,
        submitted: &HashMap<String, Array1<T>>,
        filtered: &HashMap<String, Array1<T>>,
        anomaly_results: &HashMap<String, AnomalyScore>,
        aggregate: &Array1<T>,
    ) -> Result<()> {
        let ordered: Vec<(String, Array1<T>)> = Self::ordered_cohort(submitted)
            .into_iter()
            .map(|(id, g)| (id.clone(), g.clone()))
            .collect();
        for (participant_id, gradient) in ordered {
            let participated = filtered.contains_key(&participant_id);
            let norm = from_scalar(l2_norm(&gradient))?;
            let similarity = from_scalar(cosine_similarity(&gradient, aggregate)?)?;
            let anomaly = anomaly_results
                .get(&participant_id)
                .map(|score| score.score.clamp(0.0, 1.0));
            let history = self.behavior_history.entry(participant_id).or_default();
            BehaviorHistory::push_bounded(&mut history.participation_pattern, participated);
            if participated {
                history.rounds_participated = history.rounds_participated.saturating_add(1);
                BehaviorHistory::push_bounded(&mut history.gradient_norms, norm);
                BehaviorHistory::push_bounded(&mut history.gradient_similarities, similarity);
                if let Some(anomaly) = anomaly {
                    BehaviorHistory::push_bounded(&mut history.anomaly_scores, anomaly);
                }
            }
        }
        Ok(())
    }
    /// Update participant reputations based on aggregation results.
    ///
    /// Reputation is an exponential moving average driven by `reputation_decay`;
    /// see [`super::types::ByzantineConfig::reputation_decay`].
    pub(super) fn update_reputations(
        &mut self,
        honest_participants: &HashMap<String, Array1<T>>,
        byzantine_participants: &[String],
    ) -> Result<()> {
        let decay = self.config.reputation_decay;
        let honest_rate = 1.0 - decay;
        let byzantine_rate = (honest_rate * 5.0).min(1.0);
        let mut honest_ids: Vec<&String> = honest_participants.keys().collect();
        honest_ids.sort();
        for participant_id in honest_ids {
            let quality = self.derived_quality(participant_id);
            let reputation = self
                .reputation_scores
                .entry(participant_id.clone())
                .or_default();
            reputation.successful_aggregations =
                reputation.successful_aggregations.saturating_add(1);
            reputation.score = (decay * reputation.score + honest_rate).clamp(0.0, 1.0);
            if let Some((gradient_quality, consistency)) = quality {
                reputation.gradient_quality = gradient_quality;
                reputation.consistency_score = consistency;
            }
            reputation.trust_level = match reputation.score {
                s if s >= 0.8 => TrustLevel::High,
                s if s >= 0.5 => TrustLevel::Medium,
                _ => TrustLevel::Low,
            };
        }
        for participant_id in byzantine_participants {
            let quality = self.derived_quality(participant_id);
            let reputation = self
                .reputation_scores
                .entry(participant_id.clone())
                .or_default();
            reputation.detected_anomalies = reputation.detected_anomalies.saturating_add(1);
            reputation.score = (reputation.score * (1.0 - byzantine_rate)).clamp(0.0, 1.0);
            if let Some((gradient_quality, consistency)) = quality {
                reputation.gradient_quality = gradient_quality;
                reputation.consistency_score = consistency;
            }
            reputation.trust_level = if reputation.score < 0.1 {
                TrustLevel::Blacklisted
            } else if reputation.score < 0.3 {
                TrustLevel::Low
            } else {
                TrustLevel::Medium
            };
        }
        Ok(())
    }
    /// Gradient quality and consistency derived from the recorded behaviour history.
    pub(super) fn derived_quality(&self, participant_id: &str) -> Option<(f64, f64)> {
        let history = self.behavior_history.get(participant_id)?;
        let quality = 1.0 - history.mean_anomaly_score()?;
        let consistency = (history.mean_similarity()? + 1.0) / 2.0;
        Some((quality.clamp(0.0, 1.0), consistency.clamp(0.0, 1.0)))
    }
    /// Feed this round's outcome to the behaviour pattern model.
    pub(super) fn learn_patterns(
        &mut self,
        honest_participants: &HashMap<String, Array1<T>>,
        filtered: &HashMap<String, Array1<T>>,
        byzantine_participants: &[String],
    ) -> Result<()> {
        let honest: Vec<Array1<T>> = Self::ordered_cohort(honest_participants)
            .into_iter()
            .map(|(_, g)| g.clone())
            .collect();
        for gradient in honest {
            self.anomaly_detector.learn_normal(&gradient)?;
        }
        let attacks: Vec<Array1<T>> = byzantine_participants
            .iter()
            .filter_map(|id| filtered.get(id).cloned())
            .collect();
        for gradient in attacks {
            self.anomaly_detector.learn_attack(&gradient)?;
        }
        Ok(())
    }
    /// Compute Euclidean distance between two gradients
    pub fn compute_euclidean_distance(&self, a: &Array1<T>, b: &Array1<T>) -> Result<T> {
        euclidean_distance(a, b)
    }
    /// Compute cosine similarity between two gradients
    pub fn compute_cosine_similarity(&self, a: &Array1<T>, b: &Array1<T>) -> Result<T> {
        cosine_similarity(a, b)
    }
    /// Get reputation updates
    pub(super) fn get_reputation_updates(&self) -> HashMap<String, ReputationScore> {
        self.reputation_scores.clone()
    }
}