Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

Machine Theory of Mind for Autonomous Cyber-Defence

T0 review · 4 major / 5 minor · reviewed 2026-08-11 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read A graph-neural-network Theory of Mind model can predict an unseen cyber-attacker's target and attack path from a few past behaviour traces.

desk verdict Competent, honest paper that overclaims the headline: tree-only experiments don't support 'arbitrary topologies,' the mental network is off, but the NTD metric and held-out results are sound. read the letter →

arxiv 2412.04367 v1 pith:MJUY4IRM submitted 2024-12-05 cs.LG cs.AIcs.MA

Luke Swaby , Matthew Stewart , Daniel Harrold , Chris Willis , Gregory Palmer This is my paper · ORCID
classification cs.LGcs.AIcs.MA
keywords theoryofmindgraphneuralnetworkscyber-defencesuccessorrepresentationWassersteindistancenetworktransportautonomouscyberoperationsinterpretability
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper argues that a Theory of Mind model built on graph neural networks can infer a cyber-attacker's goal and likely attack route from only a handful of past behavioural traces. The authors introduce GIGO-ToM, an architecture that takes the network topology as a graph and outputs, for every node, a predicted probability that it is the attacker's high-value target and a predicted successor representation (the expected path toward that target). In an abstract cyber-defence simulator, GIGO-ToM is claimed to predict the targets of unseen attacking agents with weighted F1 scores up to 0.6893 across 60 possible target nodes, and to produce attack-path predictions with a mean Network Transport Distance of 0.08 on the hardest setting. The accompanying Network Transport Distance (NTD) metric is a Wasserstein distance normalized by network diameter, making it bounded and topology-aware. The practical motivation is that defenders who know where an attack is heading, rather than merely that it is happening, can allocate defensive resources in advance.

What carries the argument

The load-bearing object is GIGO-ToM, a Theory of Mind network whose character and mental encoders use graph attention layers with pooled outputs fed into an LSTM, and whose prediction heads are themselves graph neural networks, so that predictions are made on the same graph structure as the input observation. The prediction targets are the per-node probability of being the high-value target and the successor representation, both trained with a soft-label cross-entropy loss. The companion metric is the Network Transport Distance, $\text{NTD}(P,Q,D)=\frac{1}{\max(D)}\inf_{\mu\in M(P,Q)}\int_{\mathcal{X}\times\mathcal{X}} d(i,j)\,d\mu(i,j)$, where $D$ is the matrix of pairwise shortest-path lengths and $d(i,j)$ is the shortest-path distance between nodes $i$ and $j$; dividing by the network diameter bounds the score in $[0,1]$. An optional weighting function $\mathcal{W}$ linearly combines user-selected node features, scales them to $[f,1]$, and rescales input distributions before the distance is computed.

What would settle it

Run the model on a non-tree topology (for example a mesh, ring, or realistic enterprise network) with high-value nodes on interior nodes and with entry-node positions varied across episodes; a substantial drop in weighted F1 and NTD scores relative to the tree results would show that the claimed topology-agnostic generalization does not hold.

Watch

Extended reading notes

Core claim

The central claim is that GIGO-ToM, a graph-in, graph-out Theory of Mind architecture, can accurately predict both the preferred high-value target node and the attack trajectory (successor representation) of previously unseen cyber-attacking agents across a range of network topologies. All input and output layers are implemented with graph attention layers, so the output dimensionality matches the observed graph rather than being fixed in advance. The authors report that character embeddings learned from past trajectories cluster cleanly according to the attacking agent's target preference, and that target and trajectory predictions improve when more past trajectories are supplied. Compared with GIDO-ToM, a dense-output variant that more closely resembles the original ToMnet, GIGO-ToM achieves substantially higher weighted F1 scores for target prediction and consistently lower (better) NTD scores for trajectory prediction. The NTD itself is presented as a unit-bounded, network-agnostic metric that respects graph structure and can be optionally weighted by user-selected node features.

Load-bearing premise

The claimed applicability to arbitrary network topologies rests on experiments performed only on five custom tree-shaped networks, with a single fixed entry node and high-value nodes always placed on leaf nodes.

Editorial extensions

If this is right

  • A cyber defender equipped with this model could identify which machine an attacker is heading for and which intermediate nodes are at risk, from a few prior attack episodes.
  • Providing more past trajectories to the character network improves both target and attack-path predictions, so historical attack logs become a direct source of predictive power.
  • The performance is reported to be robust to network size within the tested range, although it degrades as the number of branches (and hence plausible targets) increases.
  • The NTD can be used to compare predictions across networks of different sizes on a common 0-to-1 scale, and its weighting function lets operators focus the evaluation on nodes that matter strategically.
  • Initial experiments reported in an appendix indicate that a Sinkhorn-based, differentiable version of the NTD can serve as a training loss and improves successor-representation predictions relative to cross-entropy.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The generalization claim is likely to be sensitive to input distribution: real networks with interconnections, varied entry points, or high-value nodes on interior nodes are outside the evaluated regime, so the reported accuracy should not be assumed to carry over without retesting.
  • Because the model occasionally hedges by assigning probability to paths toward multiple targets (roughly 20% of the hardest samples), a practical defender could treat such outputs as an explicit uncertainty signal and plan for several possible attack routes.
  • The NTD loss could be exported to other graph-structured prediction tasks, such as traffic-flow prediction or logistics routing, wherever a bounded, topology-aware distance between predicted and actual paths is desired.
  • A false-belief test (e.g., a cyber-domain analogue of the Sally-Anne task) would help determine whether the model exhibits genuine theory of mind or exploits statistical shortcuts in the training distribution, a distinction the paper itself flags as open.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper introduces GIGO-ToM, a graph neural network-based Theory of Mind architecture for cyber defence, and the Network Transport Distance (NTD), a diameter-normalized Wasserstein metric for evaluating predicted successor representations on graphs. The model is trained on a custom hot-desking user problem in the Yawning-Titan environment and evaluated on five tree topologies against a dense-output baseline (GIDO-ToM). The central claims are that GIGO-ToM accurately predicts the high-value target node and attack trajectory of red agents, that its learned character embeddings characterize agent policies, and that the NTD provides a bounded, topology-aware evaluation metric with an optional feature-based weighting function.

Significance. If the claims hold, the work would contribute to interpretable cyber defence by offering a GNN-based approach to anticipate attacker goals and routes from past behaviour. The NTD normalization is mathematically sound: the Wasserstein distance in a shortest-path metric space is indeed bounded by the graph diameter, and the weighting scheme is a reasonable tool for prioritising prediction errors by node features. The paper also ships a custom environment, multiple rule-based agents, and a clear held-out evaluation protocol. However, the breadth of the claims substantially exceeds the evidence: all experiments use only tree topologies from one generator with a fixed entry node and leaf-only high-value nodes, and the paper itself states that the mental network was deactivated. The empirical core is solid within its narrow setting but does not support the abstract's 'arbitrary computer network topologies' claim or the full 'theory of mind' framing.

major comments (4)
  1. [Abstract; Experiments, 'Games'] The claim of prediction 'over arbitrary computer network topologies' is not supported by the experiments. All training and evaluation are restricted to the TreeNetworkMixed setting, which consists of five custom tree topologies (TreeNetwork30, 40, 50, 70, 90) generated by the authors' own TreeNetwork generator, with a single fixed entry node and high-value nodes always placed on leaves. The two additional topologies described in Appendix A (ForestNetwork and OpticalCoreNetwork) are not used for the ToM experiments; notably OpticalCoreNetwork places high-value nodes on central servers rather than leaves, which would be a meaningful test of the claimed transfer. The abstract should be revised to state the actual scope (tree topologies from one generator under these placement constraints), or additional experiments on non-tree, cyclic, or internal-high-value-node topologies should be reported.
  2. [Discussion, paragraph beginning 'We note that there are numerous avenues'] The paper states that the mental network 'was deactivated for our experiments.' This means the model conditions only on past trajectories through the character network and does not use the current episode's observations or an inferred mental state, so the architecture is not a full ToMnet as defined in Section 'Machine Theory of Mind' and Figure 1. The abstract's characterization of the approach as 'Theory of Mind' and its claims about predicting 'contextual beliefs' are therefore misleading. Either the experiments should be run with the mental network active, or the claims should be explicitly restricted to behavior prediction from past history rather than mental-state inference.
  3. [Experiments, 'Games'; Abstract] The abstract claims prediction of 'various unseen cyber-attacking agents,' but the only red agents used in the ToM experiments are 1,000 parameterizations of a single rule-based species, RedHVTPreferenceSP, with preferences sampled from a Dirichlet distribution with concentration parameter α = 0.01. The other red agent types described in Appendix B (e.g., RedRandomSimple, RedTargetConnected, RedHVTSimple) are not evaluated in the ToM experiments, so the evidence does not support the 'various agents' claim. The language should be narrowed to 'agents of a single rule-based species with varied preference vectors' or additional agent types should be included.
  4. [Experiments, 'How well can GIDO-ToM/GIGO-ToM characterize various cyber-attacking agents?'; Figure 9] The claim in the abstract and this section that GIGO-ToM 'learn embeddings that can effectively characterize their policies' is supported only by qualitative t-SNE visualizations. No quantitative clustering metric (e.g., cluster purity, adjusted Rand index, or classification accuracy on held-out agents) is reported. The visual plots appear to show separation, but the strength of the claim is not commensurate with the evidence presented.
minor comments (5)
  1. [Throughout] There are numerous typographical errors, including 'abality' (Experiments), 'network-agnistic' (Introduction), 'archtecture' (Contributions Summary), 'determinstic' (Experiments and Discussion), 'comitting' (Discussion), 'it's bets' (Discussion), 'NT D' (Figure 13b caption), and 'precitive' (Introduction). A thorough proofread is needed.
  2. [Experiments, Figure 12 caption] The caption for Figure 12a reads 'Figure 10a: Mean test set attack path prediction' but should refer to Figure 12a. The same mislabeling appears to affect Figure 12b's caption.
  3. [Methods, Equations (12)-(13)] The notation for the min-max scaling function, written as ∥·∥_f, is confusing because it is not a norm. The explanation in the text is clear, but the symbol should be replaced with a non-norm notation (e.g., scale_f(·)) to avoid implying a norm.
  4. [The Network Transport Distance, after Equation (11)] The NTD is described as 'network-agnostic' and 'standardized' for comparing networks of different sizes. Dividing by the diameter makes the metric unit-bounded in [0,1], but it does not make scores directly comparable across networks: the same absolute error in a small-diameter network can produce a much larger NTD than in a large-diameter network. The text should say 'unit-bounded' rather than 'network-agnostic' to avoid overstating the property.
  5. [Data availability] The framework and data are proprietary to BAE Systems and not publicly available. This limits reproducibility of the empirical results. At minimum, the authors should provide pseudocode for the environment, the agent policies, and the exact train/validation split procedure, or state explicitly what code will be released.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the prediction results are supervised against held-out ground truth, and the NTD is an evaluation metric with user-set weights rather than a fitted input; the overbroad topology claim is a generalization gap, not a circular derivation.

full rationale

The derivation chain is self-contained. GIGO-ToM is trained with a weighted binary cross-entropy loss for target prediction and a soft-label cross-entropy loss for successor representations, where the ground-truth SR is obtained from empirical discounted rollouts of the observed Red agent's actual future path; predictions are then assessed on a hold-out set of 200 unseen Red agents. The Network Transport Distance is a defined rescaling of the Wasserstein distance by the graph diameter, used only to score predicted versus true successor representations; its weighting coefficients and floor parameter are user-specified, not fitted to the data, so no fitted parameter is renamed as a prediction. The author-overlapping citation (ref. 13) motivates graph representations for variable-sized inputs but does not supply the central empirical results, which stand on the paper's own train/validation split and held-out evaluations. The acknowledged simplification that entry nodes are fixed and high-value nodes are always leaves, together with the fact that all ToM experiments use five tree topologies, undermines the abstract's claim of 'arbitrary computer network topologies' as an external-validity or overclaim concern, not as circularity. The deactivation of the mental network is also a stated limitation of the ToM interpretation, but it does not make the target or trajectory predictions equivalent to their inputs. The preliminary Appendix C experiment that optimizes with an NTD-based loss and evaluates with NTD is an aligned-objective pilot, not a central claim whose inputs are defined in terms of its outputs. No step in the paper's derivation reduces, by construction or by self-citation, to its own inputs.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central results rest on a controlled distribution of rule-based agents, tree topologies, and standard optimal-transport math. The only hand-chosen parameters are the NTD weighting and the hedging classifier's k; no new physical entities are introduced.

free parameters (3)
  • NTD weighting coefficients c_i and floor f = f = 0.1; c = -1, 0, +1 in experiments
    User-specified parameters of weighting function W (Eq. 12-13) that reweight node probabilities before computing NTDθ; varying them changes reported scores and the spatial conclusions in Figure 13.
  • K-means cluster count k = 4
    Chosen post hoc to classify predicted attack paths into 3 high-value-node branches plus a 'miscellaneous' bin; the claim that ~20% of hardest predictions are hedged depends on this choice (Figure 15).
  • Dirichlet concentration α for Red agent preferences = 0.01
    Chosen to generate sparse preference vectors π for the RedHVTPreferenceSP agent species; defines the distribution of agent policies the model must distinguish.
assumptions (5)
  • standard math Wasserstein distance with shortest-path ground metric is bounded by the graph diameter, so NTD lies in [0,1].
    Used to define the normalization in Eq. 9 and the equivalence in Eq. 11.
  • domain assumption Each Red agent has a fixed preference vector over high-value nodes and always takes the shortest path to its chosen target (RedHVTPreferenceSP).
    Makes the target and trajectory predictable from past episodes; stated in Agents section.
  • domain assumption High-value nodes are always leaf nodes and the entry node is fixed across episodes.
    Simplifies target prediction and limits topology coverage; stated in Experiments, Games paragraph.
  • domain assumption Only episodes that end in Red reaching the high-value node are used for trajectory evaluation.
    Selection criterion for SR evaluation; stated in attack trajectory results section.
  • ad hoc to paper The character embedding learned from Npast past trajectories is sufficient to predict the agent's goal and path without using the mental network.
    The mental network was deactivated for all experiments (Discussion), so the model's predictive power rests on this sufficiency assumption.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Machine Theory of Mind for Autonomous Cyber-Defence." pith.science (2026). https://pith.science/paper/MJUY4IRM

@misc{pith2026241204367,
  author       = {Pith},
  title        = {Pith review of: Machine Theory of Mind for Autonomous Cyber-Defence},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MJUY4IRM}},
  note         = {Machine review of arXiv:2412.04367}
}
read the original abstract

Intelligent autonomous agents hold much potential for the domain of cyber-security. However, due to many state-of-the-art approaches relying on uninterpretable black-box models, there is growing demand for methods that offer stakeholders clear and actionable insights into their latent beliefs and motivations. To address this, we evaluate Theory of Mind (ToM) approaches for Autonomous Cyber Operations. Upon learning a robust prior, ToM models can predict an agent's goals, behaviours, and contextual beliefs given only a handful of past behaviour observations. In this paper, we introduce a novel Graph Neural Network (GNN)-based ToM architecture tailored for cyber-defence, Graph-In, Graph-Out (GIGO)-ToM, which can accurately predict both the targets and attack trajectories of adversarial cyber agents over arbitrary computer network topologies. To evaluate the latter, we propose a novel extension of the Wasserstein distance for measuring the similarity of graph-based probability distributions. Whereas the standard Wasserstein distance lacks a fixed reference scale, we introduce a graph-theoretic normalization factor that enables a standardized comparison between networks of different sizes. We furnish this metric, which we term the Network Transport Distance (NTD), with a weighting function that emphasizes predictions according to custom node features, allowing network operators to explore arbitrary strategic considerations. Benchmarked against a Graph-In, Dense-Out (GIDO)-ToM architecture in an abstract cyber-defence environment, our empirical evaluations show that GIGO-ToM can accurately predict the goals and behaviours of various unseen cyber-attacking agents across a range of network topologies, as well as learn embeddings that can effectively characterize their policies.

Figures

Figures reproduced from arXiv: 2412.04367 by the authors.

Figure 1
Figure 1. The original ToMnet architecture, adopted from Rabinowitz et al.33 It consists of three components: a character network, a mental network, and a prediction network. The character network parses observations of past behaviour (e.g. previous episodes) to build a character embedding echar. The mental network then uses this along with behavioural observations from a current episode to build a representation of the agent… view at source ↗
Figure 2
Figure 2. An illustration of the hot-desking user problem for cyber-defence. The node labelled ‘E’ denotes the entry node for the Red cyber-attacking agent. Blue, red, and pink nodes represent current user locations. The color weighting of the remaining nodes represents their vulnerability to attacks, with darker shades representing higher vulnerability scores. In each episode the users may relocate to a different desk. There… view at source ↗
Figure 3
Figure 3. Example visualizations of the five custom YAWNING-TITAN TreeNetwork topologies used in our experiments. For each topology, the node labelled ‘E’ denotes the entry node for the Red agent. Pink nodes represent high-value nodes, with the dark blue node being the one ultimately targeted by a hypothetical Red agent. Agents To enable an extensive evaluation, we have additionally implemented a range of rule-based Blue cybe… view at source ↗
Figures from the paper (14 more)
Figure 4
Figure 4. Figure 4: Shared architectural components of GIGO-ToM and GIDO-ToM. Graph-In, Dense-Out ToMnet Similar to the original ToMnet, for GIDO-ToM, the character and mental embeddings, ec and em, along with the current state observation s (obs) t , are fed into a prediction network. Th…
Figure 5
Figure 5. Figure 5: The Graph-In Dense-Out (GIDO-ToM) prediction network architecture. Output layers are implemented using dense neural network layers, necessitating a fixed number of output nodes. Graph-In, Graph-Out ToMnet In contrast to GIDO-ToM, for GIGO-ToM we concatenate the charact…
Figure 6
Figure 6. Figure 6: The Graph-In Graph-Out (GIGO-ToM) prediction network architecture. Both feature extractors and output layers are implemented using graph neural network layers, providing a flexible formulation that can parse inputs and generate outputs of variable dimensions. The Netwo…
Figure 7
Figure 7. Figure 7: Hypothetical example demonstrating the importance of a topology-aware metric for evaluating graph-based successor representations. Shown is a 50-node network with three distinct attack paths overlaid: a ground-truth t¯, and two predictions p¯, q¯. For demonstrative pur…
Figure 8
Figure 8. Figure 8: An illustrative example showing the influence of our weighting function W on output NTDθ scores under different node feature weighting parameterizations. The top figures display a ground truth and predicted attack path (SRˆ true and SRˆ pred) for TreeNetwork30. Darker …
Figure 9
Figure 9. Figure 9: shows clustered character network embeddings echar for both GIGO-ToM and our benchmark model, GIDO-ToM, over hold-out test sets of unseen Red agents when given access to varing numbers of past behaviour observations (Npast ∈ {1,2,3,4}). We find that while GIDO-ToM gene…
Figure 10
Figure 10. Figure 10: Figure 10a: high-value node prediction F1 scores for both GIGO-ToM and GIDO-ToM over the entire test set. Results are recorded for runs across different values of Npast (number of past trajectories presented to the character network). Figure 10b: Stratified F1 scores …
Figure 11
Figure 11. Figure 11: GIDO-ToM and GIGO-ToM confusion matrices for high-value node prediction across different values of Npast (number of past trajectories presented to the character network) for TreeNetworkMixed. For each confusion matrix, columns represent predicted target nodes and rows…
Figure 12
Figure 12. Figure 12: Figure 10a: Mean test set attack path prediction (SRˆ pred) NTD scores for both GIGO-ToM and GIDO-ToM across different values of Npast (number of past trajectories presented to the character network). Figure 10b: Violin plot with overlaid box plot displaying test set …
Figure 13
Figure 13. Figure 13: Figure 13a: Violin plot with overlaid box plot displaying our spatio-temporal evaluation of predicted test set attack paths: Weighted NTDθ scores (GIGO-ToM, TreeNetworkMixed, Npast = 4), stratified by node remoteness weighting (spatial) and discount factor γ (temporal…
Figure 14
Figure 14. Figure 14: Evidence of GIGO-ToM’s ‘hedging’ behaviour: SRˆ true vs. SRˆ pred for the test sample with the largest discrepancy between NT D−1 and NT D1 (i.e. for which the node remoteness weighting produces the greatest difference in NTDθ score). vˆtrue and ˆvpred are emboldened …
Figure 15
Figure 15. Figure 15: Histogram displaying the distribution of test samples for the TreeNetwork90, γ = 0.999 condition across different k-means groupings. K-means over the predicted successor representations for these samples with k = 4 (3 high-value nodes plus a ‘miscellaneous’ bin) expos…
Figure 16
Figure 16. Figure 16: Additional YAWNING-TITAN Network Topologies B Rule-Based Blue Cyber-Defence and Red Attacking Agents In addition to the agents used for the experiments in this report, we initially developed a suite of other Blue cyber-defence and Red cyber-attacking agents for our ev…
Figure 17
Figure 17. Figure 17: Evaluation and test set NTD scores for GIGO-ToM’s predicted successor representations over TreeNetworkMixed when training GIGO-ToM only to optimize SR predictions using our NTD loss function, benchmarked against the soft label cross-entropy loss (SXE). Mean NTD scores…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. One Model, Two Minds: A Context-Gated Graph Learner that Recreates Human Biases

    cs.AI 2025-09 reject novelty 4.0 of 10

    A graph-based dual-process model with a learned gate claims to reproduce four cognitive biases in theory-of-mind tasks, but the bias effects are mostly learned from supervised labels rather than emergent.

Reference graph

Works this paper leans on

60 extracted references · 56 canonical work pages · cited by 1 Pith paper

  1. [1]

    Pendleton, S. D. et al. Perception, planning, control, and coordination for autonomous vehicles (2017)

    work page 2017

  2. [2]

    Army of none: Autonomous weapons and the future of war (WW Norton & Company, 2018)

    Scharre, P. Army of none: Autonomous weapons and the future of war (WW Norton & Company, 2018)

    work page 2018

  3. [3]

    Yang, G.-Z. et al. Medical robotics—regulatory, ethical, and legal considerations for increasing levels of autonomy (2017)

    work page 2017

  4. [4]

    Benitti, F. B. V . Exploring the educational potential of robotics in schools: A systematic review (2012)

    work page 2012

  5. [5]

    & Chakraborty, S

    Preece, A., Harborne, D., Braines, D., Tomsett, R. & Chakraborty, S. Stakeholders in explainable ai (2018)

    work page 2018

  6. [6]

    P., Brundage, M

    Arulkumaran, K., Deisenroth, M. P., Brundage, M. & Bharath, A. A. Deep reinforcement learning: A brief survey, DOI: 10.1109/msp.2017.2743240 (2017)

    arXiv 2017

  7. [7]

    Perception and navigation in autonomous systems in the era of learning: A survey (2022)

    Tang, Y .et al. Perception and navigation in autonomous systems in the era of learning: A survey (2022)

    work page 2022

  8. [8]

    Deep learning-based autonomous driving systems: A survey of attacks and defenses (2021)

    Deng, Y .et al. Deep learning-based autonomous driving systems: A survey of attacks and defenses (2021). Unpublished Work Copyright 2024 BAE Systems. All Rights Reserved. 18/29

    work page 2021

Show all 60 references
  1. [9]

    Adawadkar, A. M. K. & Kulkarni, N. Cyber-security and reinforcement learning—a brief survey (2022)

    2022

  2. [10]

    & Qiu, M

    Li, C. & Qiu, M. Reinforcement learning for cyber-physical systems: with cybersecurity case studies (Chapman and Hall/CRC, 2019)

    2019

  3. [11]

    Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead (2019)

    Rudin, C. Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead (2019)

    2019

  4. [12]

    Du, M., Liu, N. & Hu, X. Techniques for interpretable machine learning (2019). 1808.00033

    2019 arXiv

  5. [13]

    Palmer, G., Parry, C., Harrold, D. J. & Willis, C. Deep reinforcement learning for autonomous cyber operations: A survey (2023)

    2023

  6. [14]

    Miles, I. et al. Reinforcement learning for autonomous resilient cyber defence (2024)

    2024

  7. [15]

    Deng, S. et al. Edge intelligence: The confluence of edge computing and artificial intelligence (2020)

    2020

  8. [16]

    The near-term impact of ai on the cyber threat (2024)

    NCSC. The near-term impact of ai on the cyber threat (2024). Accessed: 25/01/2024

    2024

  9. [17]

    Nisioti, A., Mylonas, A., Yoo, P. D. & Katos, V . From intrusion detection to attacker attribution: A comprehensive survey of unsupervised methods (2018)

    2018

  10. [18]

    Kala, E. S. M. Critical role of cyber security in global economy (2023)

    2023

  11. [19]

    F., Commission, A

    Police, A. F., Commission, A. C. I. et al. Acsc annual cyber threat report: July 2019 to june 2020 (2020)

    2020

  12. [20]

    Moustafa, N., Koroniotis, N., Keshk, M., Zomaya, A. Y . & Tari, Z. Explainable intrusion detection for cyber defences in the internet of things: Opportunities and solutions (2023)

    2023

  13. [21]

    & Fern, A

    Greydanus, S., Koul, A., Dodge, J. & Fern, A. Visualizing and understanding atari agents. In International conference on machine learning, 1792–1801 (PMLR, 2018)

    2018

  14. [22]

    Robust artificial intelligence for active cyber defence (2020)

    Burke, A. Robust artificial intelligence for active cyber defence (2020)

    2020

  15. [23]

    Antoniadi, A. M. et al. Current challenges and future opportunities for xai in machine learning-based clinical decision support systems: a systematic review (2021)

    2021

  16. [24]

    Explanation in artificial intelligence: Insights from the social sciences (2019)

    Miller, T. Explanation in artificial intelligence: Insights from the social sciences (2019)

    2019

  17. [25]

    Lipton, Z. C. The mythos of model interpretability (2018)

    2018

  18. [26]

    Samek, W., Montavon, G., Vedaldi, A., Hansen, L. K. & Müller, K.-R. Explainable AI: interpreting, explaining and visualizing deep learning, vol. 11700 (Springer Nature, 2019)

    2019

  19. [27]

    & Woodruff, G

    Premack, D. & Woodruff, G. Does the chimpanzee have a theory of mind? (1978)

    1978

  20. [28]

    theory of mind

    Baron-Cohen, S., Leslie, A. M. & Frith, U. Does the autistic child have a “theory of mind”? (1985)

    1985

  21. [29]

    L., Peterson, C

    Garfield, J. L., Peterson, C. C. & Perry, T. Social cognition, language acquisition and the development of the theory of mind (2001)

    2001

  22. [30]

    Hamilton, A. F. d. C. Research review: Goals, intentions and mental states: Challenges for theories of autism (2009)

    2009

  23. [31]

    Ahmed, F. S. & Stephen Miller, L. Executive function mechanisms of theory of mind (2011)

    2011

  24. [32]

    & Stich, S

    Nichols, S. & Stich, S. P. Mindreading: An integrated account of pretence, self-awareness, and understanding other minds (Oxford University Press, 2003)

    2003

  25. [33]

    Rabinowitz, N. et al. Machine theory of mind. In International conference on machine learning , 4218–4227 (PMLR, 2018)

    2018

  26. [34]

    Improving generalization for temporal difference learning: The successor representation (1993)

    Dayan, P. Improving generalization for temporal difference learning: The successor representation (1993)

    1993

  27. [35]

    & Dhir, N

    Andrew, A., Spillard, S., Collyer, J. & Dhir, N. Developing optimal causal cyber-defence agents via cyber security simulation. In Workshop on Machine Learning for Cybersecurity (ML4Cyber) (2022)

    2022

  28. [36]

    & Guibas, L

    Rubner, Y ., Tomasi, C. & Guibas, L. J. A metric for distributions with applications to image databases. InSixth international conference on computer vision (IEEE Cat. No. 98CH36271) , 59–66 (IEEE, 1998)

    1998

  29. [37]

    & Wang, Y

    Wang, Y ., Zhong, F., Xu, J. & Wang, Y . Tom2c: Target-oriented multi-agent communication and cooperation with theory of mind (2021)

    2021

  30. [38]

    Shu, T. et al. Agent: A benchmark for core psychological reasoning. In International Conference on Machine Learning , 9614–9625 (PMLR, 2021). Unpublished Work Copyright 2024 BAE Systems. All Rights Reserved. 19/29

    2021

  31. [39]

    & Kim, K.-J

    Jeon, H., Oh, S., You, W., Jung, H. & Kim, K.-J. Inferring relationship using theory of mind in press diplomacy. In ICML 2022 Workshop AI for Agent-Based Modelling (2022)

    2022

  32. [40]

    & Behzadan, V

    Piazza, N. & Behzadan, V . A theory of mind approach as test-time mitigation against emergent adversarial communication (2023)

    2023

  33. [41]

    & Zhou, B

    Cheng, Q., Wu, C., Hu, B., Kong, D. & Zhou, B. Think that attackers think: Using first-order theory of mind in intrusion response system. In 2019 IEEE Global Communications Conference (GLOBECOM) , 1–6, DOI: 10.1109/ GLOBECOM38437.2019.9013291 (2019)

    2019

  34. [42]

    & Gonzalez, C

    Malloy, T. & Gonzalez, C. Learning to defend by attacking (and vice-versa): Transfer of learning in cybersecurity games. In 2023 IEEE European Symposium on Security and Privacy Workshops (EuroS&PW) , 458–464, DOI: 10.1109/ EuroSPW59978.2023.00056 (2023)

    2023

  35. [43]

    V on Stackelberg, H.Market structure and equilibrium (Springer Science & Business Media, 2010)

    2010

  36. [44]

    Wang, J., Xu, C., Yang, W. & Yu, L. A normalized gaussian wasserstein distance for tiny object detection (2021)

    2021

  37. [45]

    H., Yun, I

    Kim, D. H., Yun, I. D. & Lee, S. U. A new attributed relational graph matching algorithm using the nested structure of earth mover’s distance. In Proceedings of the 17th International Conference on Pattern Recognition, 2004. ICPR 2004. , vol. 1, 48–51 (IEEE, 2004)

    2004

  38. [46]

    & De Bie, T

    Noels, S., Vandermarliere, B., Bastiaensen, K. & De Bie, T. An earth mover’s distance based graph distance metric for financial statements. In 2022 IEEE Symposium on Computational Intelligence for Financial Engineering and Economics (CIFER), 1–8 (IEEE, 2022)

    2022

  39. [47]

    & Stehn, F

    Cheong, O., Gudmundsson, J., Kim, H.-S., Schymura, D. & Stehn, F. Measuring the similarity of geometric graphs. In Experimental Algorithms: 8th International Symposium, SEA 2009, Dortmund, Germany, June 4-6, 2009. Proceedings 8 , 101–112 (Springer, 2009)

    2009

  40. [48]

    Graph mover’s distance: An efficiently computable distance measure for geometric graphs (2023)

    Majhi, S. Graph mover’s distance: An efficiently computable distance measure for geometric graphs (2023)

    2023

  41. [49]

    Shapley, L. S. Stochastic games (1953)

    1953

  42. [50]

    & Schmidhuber, J

    Hochreiter, S. & Schmidhuber, J. Long short-term memory (1997)

    1997

  43. [51]

    Vaswani, A. et al. Attention is all you need (2017)

    2017

  44. [52]

    Veliˇckovi´c, P. et al. Graph attention networks. In International Conference on Learning Representations (2018)

    2018

  45. [53]

    & Yahav, E

    Alon, U. & Yahav, E. On the bottleneck of graph neural networks and its practical implications (2020)

    2020

  46. [54]

    & Kim, B

    Doshi-Velez, F. & Kim, B. Towards a rigorous science of interpretable machine learning (2017)

    2017

  47. [55]

    & Hinton, G

    Van der Maaten, L. & Hinton, G. Visualizing data using t-SNE. (2008)

    2008

  48. [56]

    & Perner, J

    Wimmer, H. & Perner, J. Beliefs about beliefs: Representation and constraining function of wrong beliefs in young children’s understanding of deception (1983)

    1983

  49. [57]

    Geirhos, R. et al. Shortcut learning in deep neural networks (2020)

    2020

  50. [58]

    & Sarkadi, S

    Piazza, N., Behzadan, V . & Sarkadi, S. Limitations of theory of mind defenses against deception in multi-agent systems (2023)

    2023

  51. [59]

    & Knopp, P

    Sinkhorn, R. & Knopp, P. Concerning nonnegative matrices and doubly stochastic matrices (1967)

    1967

  52. [60]

    Implementation of batched sinkhorn iterations for entropy-regularized wasserstein loss (2019)

    Viehmann, T. Implementation of batched sinkhorn iterations for entropy-regularized wasserstein loss (2019). Acknowledgements Research funded by Frazer-Nash Consultancy Ltd. on behalf of the Defence Science and Technology Laboratory (Dstl) which is an executive agency of the UK...

    2019

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.