PaperlayerView exact revision on arXiv

Source-linked AI summary

TrustFormer: Cross-Temporal and Cross- Dimensional Transformer for Task-Specific Multi-Dimensional Trust Evaluation

Botao Zhu, Xianbin Wang

arXiv:2608.25238v1eess.SYcs.LG

TL;DR

Scalar trust representations inadequately capture task-relevant collaborator reliability, while heterogeneous asynchronous data and cross-dimensional dependencies complicate multi-dimensional evaluation. TrustFormer synchronizes historical records and applies cross-temporal and cross-dimensional attention to learn task-specific trust profiles. It reports a 40.8% reduction in strategic-collaborator MSE over LSTM and 91.3% overall collaborator selection accuracy.

Abstract

from arXiv · show

In dynamic collaborative systems, the selection of reliable collaborators is critical to ensuring effective task execution. Existing trust evaluation methods often rely on unidimensional or scalar representations, which fail to faithfully capture a collaborator's true trustworthiness, thereby motivating a shift toward multi-dimensional trust modeling. However, due to the asynchrony of collected trust-related data across different dimensions, as well as the complex intra- and inter-dimensional dependencies embedded within these data, multi-dimensional trust evaluation remains challenging. To address these challenges, we propose TrustFormer, a task-specific multi-dimensional trust evaluation framework. Specifically, TrustFormer leverages task identifiers and device-generated timestamps to synchronize heterogeneous trust-related data across historical collaborations. It further employs cross-temporal and cross-dimensional attention mechanisms to jointly model temporal dynamics and inter-dimensional correlations, thereby effectively learning the multi-dimensional trust evolution of potential collaborators from historical performance data. In addition, according to the multi-dimensional resource requirements of tasks, potential collaborators' multi-dimensional resource trust is evaluated. Finally, by synthesizing these multi-dimensional trust profiles, the framework enables the optimal collaborator selection. Experimental results demonstrate that TrustFormer outperforms existing methods by yielding a 40.8% improvement in trust evaluation accuracy and enabling more reliable collaborator selection.

I. INTRODUCTION

TrustFormer addresses the limits of scalar trust evaluation by modeling task-specific, multi-dimensional collaborator trust. It synchronizes heterogeneous historical data and uses cross-temporal and cross-dimensional attention to capture trust evolution and inter-dimensional dependencies.

  • Scalar trust aggregation can obscure dimension-specific deficiencies and make unreliable collaborator aspects difficult to identify.Source: §I · p. 1
  • Task requirements are inherently multi-dimensional, so collaborator trustworthiness must be evaluated independently across relevant dimensions.Source: §I · p. 1
  • Asynchronous historical data from different participants must be synchronized despite device clock drift and network delays.Source: §I · pp. 1–2
  • Trust dynamics can evolve through gradual degradation, abrupt strategic shifts, or intermittent fluctuations under unstable resource conditions.Source: §I · p. 2
  • TrustFormer combines lightweight synchronization with task-specific multi-dimensional trust evaluation in a Transformer-based framework.Source: §I · p. 2
  • Cross-temporal and cross-dimensional attention models temporal evolution within dimensions and dependencies across dimensions.Source: §I · p. 2

II. SYSTEM MODEL AND PROBLEM FORMULATION

The system models collaboration among resource-constrained devices coordinated by a trusted server. Tasks specify multi-dimensional historical-trust and physical-resource requirements that guide collaborator evaluation and selection.

II. SYSTEM MODEL AND PROBLEM FORMULATION

  • The collaborative computing system contains devices that can act as task owners or collaborators, with a trust server coordinating records and selection.Source: §II · p. 2

A. Task Model with Multi-Dimensional Requirements

B. Task-Specific Multi-Dimensional Trust Model

TrustFormer represents each candidate collaborator through task-specific historical and resource trust profiles. These profiles capture performance across multiple dimensions and whether physical resource requirements are satisfied.

C. Problem Formulation

The paper formulates collaborator selection as task-specific optimization over historical and resource trust, while TrustFormer synchronizes and learns multi-dimensional trust from heterogeneous collaboration records. Data are linked by task identifiers and bounded using device-generated timestamps.

C. Problem Formulation

  • The trust server seeks an optimal collaborator that maximizes aggregate trust while satisfying the task’s multi-dimensional requirements.Source: C. Problem Formulation · p. 3
  • Uniform weighting across trust dimensions is represented by an all-ones vector, while task-specific weights can replace it.Source: C. Problem Formulation · p. 3

III. TRUSTFORMER FOR MULTI-DIMENSIONAL TRUST EVALUATION

  • TrustFormer synchronizes historical performance data and jointly models temporal dynamics within dimensions and interdependencies across dimensions.Source: §III · p. 3

A. Multi-Dimensional Historical Data Collection

B. Multi-Dimensional Historical Data Synchronization

TrustFormer synchronizes heterogeneous historical performance data using local timestamps and collaboration-session grouping rather than requiring a unified global clock. It then derives CPU and memory stability metrics from chronologically ordered monitoring sequences.

Synchronization

  • TrustFormer groups received records into collaboration sessions and preserves within-dimension chronological order without requiring strict global-clock synchronization.Local timestamps and task-related organization support localized temporal alignment despite network delays and unsynchronized device clocks.Source: B. Multi-Dimensional Historical Data… · pp. 3–4
  • Records from the same physical device are sorted by embedded local timestamps to reconstruct the chronological sequence of each performance dimension.The method relies on the device's correct relative execution order.Source: B. Multi-Dimensional Historical Data… · p. 4

Stability metrics

  • CPU and memory utilization sequences are aggregated into scalar stability values using the root mean square of successive differences.This preserves temporal variation that simple summary statistics would discard.Source: B. Multi-Dimensional Historical Data… · p. 4
  • The resulting historical performance vector contains task completion accuracy, timeliness, CPU stability, and memory stability.The vector is associated with the server-side chronological records for the collaborator's task.Source: B. Multi-Dimensional Historical Data… · p. 4

C. Data Preprocessing when New Tasks Arrive

When a new task arrives, the trust server retrieves each collaborator's completed collaborations and constructs a chronologically ordered matrix with matching timestamps. Rows represent historical collaborations, while columns trace individual performance dimensions.

Historical data assembly

  • For each potential collaborator, the server retrieves performance vectors from K previously completed collaborations after a new task request arrives.These vectors are stacked in strictly chronological order together with their associated timestamps.Source: C. Data Preprocessing when New Tasks… · p. 4
  • The temporal performance matrix Xao is formed by vertically concatenating historical performance vectors row by row.The matrix has shape R^K×|Dhis|.Source: C. Data Preprocessing when New Tasks… · p. 4

Temporal organization

D. Multi-Dimensional Trust Representation Learning via Cross-Temporal and Cross-Dimensional Attentions

TrustFormer uses embeddings and two attention stages to model both how each trust dimension changes over time and how dimensions correlate within each collaboration. Stacked attention blocks produce representations encoding both dependencies.

Embedding

  • Each historical dimension value is projected through a dimension-specific linear layer, while continuous time intervals and positional embeddings provide temporal information.The resulting temporal embedding is shared across dimensions within the same historical record.Source: D. Multi-Dimensional Trust… · p. 4

Cross-temporal attention

  • Cross-temporal attention applies multi-head self-attention independently across historical collaborations for each performance dimension.It can identify gradual degradation and abrupt behavioral changes.Source: D. Multi-Dimensional Trust… · pp. 4–5

Cross-dimensional attention

  • Cross-dimensional attention applies multi-head self-attention across performance dimensions within each collaboration.Its attention matrix captures pairwise dependencies among dimensions.Source: D. Multi-Dimensional Trust… · p. 5

Integrated representation

  • The combined attention stages form a time-dimension block, and stacking L blocks yields an output encoding temporal dynamics and inter-dimensional dependencies.The output has shape R^K×|Dhis|×dmodel.Source: D. Multi-Dimensional Trust… · pp. 4–5

E. Historical Trust Calculation

Historical trust is computed by temporally pooling each performance dimension and passing the fused representations through an MLP with sigmoid activation. This produces the collaborator's multi-dimensional historical trust evaluation.

Historical trust computation

  • TrustFormer averages each dimension's representation along the temporal axis to obtain a temporally fused representation.The fused representation for each dimension is then passed through an MLP.Source: E. Historical Trust Calculation · p. 5
  • A sigmoid activation is used in the MLP-based computation of the multi-dimensional historical trust evaluation.The evaluation is produced separately across the historical trust dimensions.Source: E. Historical Trust Calculation · p. 5

IV. RESOURCE EVALUATION AND COLLABORATOR SELECTION

TrustFormer evaluates each potential collaborator’s resource trust dimension by dimension against the task’s required capabilities, producing a multi-dimensional resource profile.

  • Resource trust is set to 1 when a reported capability satisfies the corresponding task requirement and 0 otherwise.The server applies this evaluation separately for every required resource dimension.Source: §IV · p. 5

V. RESULT ANALYSIS

The experiments model heterogeneous device behaviors and evaluate TrustFormer under temporal disorder and diverse collaborator dynamics. TrustFormer is especially effective for abrupt strategic shifts and gradual degradation, while synchronization materially affects accuracy.

Experimental Setup

  • 20,000 collaboration records are split chronologically into 70% training, 10% validation, and 20% test sets.The simulation uses 200 virtual devices with behavior proportions of 60% reliable, 10% low-capability, 10% degradation, 10% intermittent anomaly, and 10% strategic cheating.Source: §V · pp. 5–6

Impact of Synchronization

Behavior-Type Accuracy

  • 40.8% lower MSE than LSTM and 28.3% lower than standard Transformer are achieved for strategic collaborators.For degradation collaborators, TrustFormer reduces MSE by 37.5% and 22.1% against the two baselines, respectively.Source: B. Trust Evaluation Accuracy Across… · p. 6
  • Comparable accuracy is observed across methods for reliable collaborators, while TrustFormer’s advantage widens for complex behavioral dynamics.The comparison covers reliable, low-capability, degradation, intermittent anomaly, and strategic collaborator behaviors through the simulation setup.Source: §V · p. 5, B. Trust Evaluation Accuracy Across… · p. 6

C. Collaborator Selection Accuracy Comparison

TrustFormer is evaluated for collaborator selection across task requirement dimensions and overall simultaneous satisfaction. It consistently outperforms LSTM and Transformer, with the largest reported advantage in overall accuracy.

Collaborator Selection Accuracy Comparison

  • 91.3% overall collaborator selection accuracy is achieved by TrustFormer, compared with 85.4% for Transformer and 80.7% for LSTM.Overall accuracy requires all dimensions to be satisfied simultaneously.Source: C. Collaborator Selection Accuracy… · p. 6
  • TrustFormer consistently outperforms LSTM and Transformer across all evaluated dimensions.The experiment reports collaborator selection accuracy at both per-dimension and overall levels.Source: C. Collaborator Selection Accuracy… · p. 6, Figure 2 · p. 6
  • More accurate multi-dimensional trust evaluation translates into more reliable collaborator selection.The conclusion links TrustFormer’s trust vectors and selection reliability, with notable advantages under strategic deception and gradual degradation.Source: C. Collaborator Selection Accuracy… · p. 6, §VI · p. 6
Loading 2608.25238v1…