pith. sign in

arxiv: 1907.08377 · v1 · pith:HLKKCRJInew · submitted 2019-07-19 · 💻 cs.LG · cs.AI· cs.CR

DaiMoN: A Decentralized Artificial Intelligence Model Network

Surat Teerapittayanon , H. T. Kung This is my paper

Pith reviewed 2026-05-24 19:17 UTC · model grok-4.3

classification 💻 cs.LG cs.AIcs.CR
keywords decentralized AIproof of improvementdistance embedding for labelsmodel verificationdecentralized ledgercollaborative machine learninglabel privacyincentive mechanism
0
0 comments X

The pith

DaiMoN lets peers verify machine learning model accuracy improvements without access to the test labels.

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

DaiMoN sets up an autonomous network in which peers submit classifier models that raise accuracy on a shared classification task and other peers confirm those gains. An append-only ledger records each improvement along with the contributor, the timing, the gain amount, and a pointer to the updated model, after which cryptographic tokens are awarded. The verification step must occur without exposing the private test labels, because full label access would let submitters overfit and cheat the system. The paper introduces a learnable Distance Embedding for Labels (DEL) that maps the true label vector into a lower-dimensional space while keeping distances to any classifier's output vector approximately intact, so peers can still measure accuracy gains. Analysis and experiments indicate that the resulting accuracy estimates remain reliable and that inverting DEL to recover the original labels is difficult.

Core claim

DaiMoN maintains a decentralized ledger of model improvements and uses a learnable Distance Embedding for Labels (DEL) to enable proof-of-improvement without access to true test labels. DEL embeds the test label vector in a low-dimensional space while approximately preserving distances to a classifier's inferred label vector, allowing peers to evaluate accuracy gains. The paper supplies both analysis and empirical results showing that accuracy assessment stays accurate under DEL and that the embedding resists inversion attacks aimed at label recovery.

What carries the argument

The learnable Distance Embedding for Labels (DEL) function, which embeds a dataset's test label vector into low-dimensional space while approximately preserving distances to any classifier's output vector.

If this is right

  • Peers can accurately assess model accuracy using only the DEL embedding.
  • The append-only ledger permanently records who improved a model, when, by how much, and where the new model is stored.
  • Cryptographic tokens reward contributors for verified accuracy gains.
  • The system reduces the incentive for intentional overfitting because submitters cannot see the test labels.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same embedding approach could be adapted to regression or ranking tasks if analogous distance-preserving maps can be learned.
  • The ledger-plus-DEL pattern might support marketplaces in which any participant can propose and monetize incremental model updates.
  • The requirement that DEL be dataset-specific suggests that each new test collection would need its own trained embedding before the network can operate on it.

Load-bearing premise

DEL approximately preserves distances between the true test label vector and a model's inferred label vector while remaining hard to invert.

What would settle it

An experiment in which accuracy scores computed from DEL embeddings systematically disagree with accuracy measured on the true hidden test labels, or in which the original test labels can be recovered from the DEL output with high fidelity.

Figures

Figures reproduced from arXiv: 1907.08377 by H. T. Kung, Surat Teerapittayanon.

Figure 2
Figure 2. Figure 2: Correlation between error (%) in x with respect to xt and the distance between them in the embedding space under f. After the neural network has been trained, we evaluate how well the learned f can preserve error in a predicted label vector x inferred by the classifier [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 1
Figure 1. Figure 1: The training and testing loss as the number of epochs [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 3
Figure 3. Figure 3: The probability p as distance decreases for varying n. where n is the dimension of a vector, β is the angle between xt and a vector on the sphere, and Ix(a, b) is the regularized incomplete beta function defined as: Ix(a, b) = B(x; a, b) B(a, b) . In the above expression, B(x; a, b) is the incomplete beta function, and B(a, b) is the beta function defined as: B(x; a, b) = Z x 0 t a−1 (1 − t) b−1 dt, B(a, b… view at source ↗
Figure 4
Figure 4. Figure 4: Error (e(f −1 (yt ), xt)) in percentage as the number of epochs increases for data generated with random x and near xt. 1: procedure PROVE(M) → πP 2: g ← digest(M) 3: y ← f(M(Z)) 4: return {g, y, pkP }skP vector xt. PoI is characterized by the PROVE and VERIFY procedures shown. As a part of the system setup, a prover P has a public and private key pair (pkP , skP ) and a verifier V has a public and private… view at source ↗
Figure 5
Figure 5. Figure 5: The reward function R(d, dc) as the distance d decreases for varying current best distance dc for a = 3. is the validator’s position, and Z>0 denotes the set of integers greater than zero. This factor encourages validators to compete to be the first one to submit the validation proof for the PoI proof in order to maximize the reward. Two is used as a base of the scaling factor here since P∞ s=1 2 −s = 1. D… view at source ↗
read the original abstract

We introduce DaiMoN, a decentralized artificial intelligence model network, which incentivizes peer collaboration in improving the accuracy of machine learning models for a given classification problem. It is an autonomous network where peers may submit models with improved accuracy and other peers may verify the accuracy improvement. The system maintains an append-only decentralized ledger to keep the log of critical information, including who has trained the model and improved its accuracy, when it has been improved, by how much it has improved, and where to find the newly updated model. DaiMoN rewards these contributing peers with cryptographic tokens. A main feature of DaiMoN is that it allows peers to verify the accuracy improvement of submitted models without knowing the test labels. This is an essential component in order to mitigate intentional model overfitting by model-improving peers. To enable this model accuracy evaluation with hidden test labels, DaiMoN uses a novel learnable Distance Embedding for Labels (DEL) function proposed in this paper. Specific to each test dataset, DEL scrambles the test label vector by embedding it in a low-dimension space while approximately preserving the distance between the dataset's test label vector and a label vector inferred by the classifier. It therefore allows proof-of-improvement (PoI) by peers without providing them access to true test labels. We provide analysis and empirical evidence that under DEL, peers can accurately assess model accuracy. We also argue that it is hard to invert the embedding function and thus, DEL is resilient against attacks aiming to recover test labels in order to cheat. Our prototype implementation of DaiMoN is available at https://github.com/steerapi/daimon.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 1 minor

Summary. The paper introduces DaiMoN, a decentralized network for collaborative improvement of ML classification models. Peers submit improved models and verify accuracy gains via a novel learnable Distance Embedding for Labels (DEL) that embeds test label vectors in low dimension while approximately preserving distances to a classifier's inferred labels, enabling proof-of-improvement without exposing true labels. An append-only decentralized ledger records contributions, improvements, and model locations; contributors are rewarded with cryptographic tokens. The abstract asserts that analysis and empirical evidence confirm DEL permits accurate accuracy assessment by peers and resists label-recovery attacks.

Significance. If the distance-preservation and inversion-resistance properties of DEL can be placed on firmer footing, DaiMoN would supply a concrete mechanism for decentralized, label-private model verification that could support blockchain-based collaborative ML. The public prototype implementation at the cited GitHub repository is a concrete asset that would allow independent reproduction.

major comments (3)
  1. [Abstract] Abstract: the central claim that 'analysis and empirical evidence' establish that peers can accurately assess model accuracy under DEL supplies neither equations defining the embedding, quantitative distortion bounds, datasets, nor error metrics; this absence directly undermines evaluation of the load-bearing distance-preservation property required for reliable proof-of-improvement.
  2. [Abstract] Abstract: the assertion that DEL 'is hard to invert' and therefore 'resilient against attacks aiming to recover test labels' is offered without a security reduction, formal hardness argument, or even the embedding dimension and training procedure; because the embedding is learned per test set, the lack of any concrete guarantee is load-bearing for the anti-cheating guarantee.
  3. [Abstract] Abstract: DEL is described as 'learnable' and 'specific to each test dataset,' yet no information is given on the training objective, the source of supervision for learning the embedding, or how the embedding is distributed to peers without leaking label information; these omissions affect both reproducibility and the claimed security properties.
minor comments (1)
  1. [Abstract] The abstract states that the ledger records 'who has trained the model' but does not clarify how authorship is authenticated or how the system prevents Sybil submissions of the same model under multiple identities.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive comments. We address each major comment below and will revise the abstract to improve accessibility to the technical details while preserving its high-level nature.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central claim that 'analysis and empirical evidence' establish that peers can accurately assess model accuracy under DEL supplies neither equations defining the embedding, quantitative distortion bounds, datasets, nor error metrics; this absence directly undermines evaluation of the load-bearing distance-preservation property required for reliable proof-of-improvement.

    Authors: The abstract is a high-level summary. The DEL embedding equations, distortion bounds, datasets, and error metrics appear in Sections 3 and 4. We will revise the abstract to reference these sections and briefly note the demonstrated properties. revision: yes

  2. Referee: [Abstract] Abstract: the assertion that DEL 'is hard to invert' and therefore 'resilient against attacks aiming to recover test labels' is offered without a security reduction, formal hardness argument, or even the embedding dimension and training procedure; because the embedding is learned per test set, the lack of any concrete guarantee is load-bearing for the anti-cheating guarantee.

    Authors: The abstract omits these specifics. Section 3 defines the embedding dimension, training procedure, and provides an argument plus empirical evidence for inversion resistance. No formal security reduction is present; we will update the abstract to state the parameters and note that resilience rests on analysis and experiments rather than a formal reduction. revision: partial

  3. Referee: [Abstract] Abstract: DEL is described as 'learnable' and 'specific to each test dataset,' yet no information is given on the training objective, the source of supervision for learning the embedding, or how the embedding is distributed to peers without leaking label information; these omissions affect both reproducibility and the claimed security properties.

    Authors: Section 3 details the distance-preserving training objective, the form of supervision, and the ledger-based distribution of embedding parameters that avoids exposing original labels. We will revise the abstract to include concise statements on these points. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected; DEL properties rest on internal analysis and experiments

full rationale

The paper introduces a novel DEL embedding as a new component and supports its distance-preservation and inversion-resistance claims via analysis and empirical evidence presented in the work itself. No equations reduce a claimed prediction to a fitted parameter by construction, no load-bearing self-citations are invoked for uniqueness or ansatz, and the verification mechanism does not rename or self-define prior results. The central claims therefore remain independent of the inputs they evaluate.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 1 invented entities

Based solely on the abstract, the central claim rests on the introduction of the DEL embedding as a new mechanism; no explicit free parameters, standard mathematical axioms, or additional invented entities beyond DEL and the ledger are stated.

invented entities (1)
  • Distance Embedding for Labels (DEL) function no independent evidence
    purpose: Scramble test label vectors into low-dimensional embeddings that approximately preserve distances for proof-of-improvement without exposing labels
    Presented as novel in the abstract; no independent evidence supplied for its properties beyond the authors' analysis claim

pith-pipeline@v0.9.0 · 5829 in / 1298 out tokens · 25478 ms · 2026-05-24T19:17:43.169425+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

  • IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear
    ?
    unclear

    Relation between the paper passage and the cited Recognition theorem.

    DEL scrambles the test label vector by embedding it in a low-dimension space while approximately preserving the distance between the dataset's test label vector and a label vector inferred by the classifier... We provide analysis and empirical evidence that under DEL, peers can accurately assess model accuracy.

  • IndisputableMonolith/Foundation/RealityFromDistinction.lean reality_from_one_distinction unclear
    ?
    unclear

    Relation between the paper passage and the cited Recognition theorem.

    Finding such a distance-preserving embedding function f is generally a challenging mathematical problem. Fortunately, we have observed empirically that we can learn this xt-specific embedding function using a neural network.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

32 extracted references · 32 canonical work pages · 5 internal anchors

  1. [1]

    Bitcoin: A peer-to-peer electronic cash system,

    S. Nakamoto, “Bitcoin: A peer-to-peer electronic cash system,” 2008

    work page 2008

  2. [2]

    Ethereum: A secure decentralised generalised trans- action ledger,

    G. Wood, “Ethereum: A secure decentralised generalised trans- action ledger,” Ethereum project yellow paper , vol. 151, pp. 1–32, 2014

    work page 2014

  3. [3]

    Very Deep Convolutional Networks for Large-Scale Image Recognition

    K. Simonyan and A. Zisserman, “Very deep convolutional networks for large-scale image recognition,” arXiv preprint arXiv:1409.1556, 2014

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  4. [4]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 770–778

    work page 2016

  5. [5]

    Going deeper with convolutions,

    C. Szegedy, W. Liu, Y . Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan, V . Vanhoucke, and A. Rabinovich, “Going deeper with convolutions,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2015, pp. 1–9

    work page 2015

  6. [6]

    BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding

    J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, “Bert: Pre-training of deep bidirectional transformers for language understanding,” arXiv preprint arXiv:1810.04805 , 2018

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  7. [7]

    Model Zoo,

    “Model Zoo,” https://github.com/BVLC/caffe/wiki/Model-Zoo

    work page

  8. [8]

    PyTorch Model Zoo,

    “PyTorch Model Zoo,” https://pytorch.org/docs/stable/ torchvision/models.html

    work page

  9. [9]

    Models and examples built with TensorFlow,

    “Models and examples built with TensorFlow,” https://github. com/tensorflow/models

    work page

  10. [10]

    Model Zoo,

    “Model Zoo,” https://modelzoo.co/

    work page

  11. [11]

    “Kaggle,” https://www.kaggle.com

    work page

  12. [12]

    Fashion-MNIST: a Novel Image Dataset for Benchmarking Machine Learning Algorithms

    H. Xiao, K. Rasul, and R. V ollgraf, “Fashion-MNIST: a novel image dataset for benchmarking machine learning algorithms,” arXiv preprint arXiv:1708.07747 , 2017

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  13. [13]

    Extensions of lipschitz mappings into a hilbert space,

    W. B. Johnson and J. Lindenstrauss, “Extensions of lipschitz mappings into a hilbert space,” Contemporary mathematics , vol. 26, no. 189-206, p. 1, 1984

    work page 1984

  14. [14]

    Optimality of the johnson- lindenstrauss lemma,

    K. G. Larsen and J. Nelson, “Optimality of the johnson- lindenstrauss lemma,” in 2017 IEEE 58th Annual Symposium on Foundations of Computer Science (FOCS) . IEEE, 2017, pp. 633–638

    work page 2017

  15. [15]

    Adam: A Method for Stochastic Optimization

    D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” arXiv preprint arXiv:1412.6980 , 2014

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  16. [16]

    Limit theorems for uniform distributions on spheres in high-dimensional euclidean spaces,

    A. Stam, “Limit theorems for uniform distributions on spheres in high-dimensional euclidean spaces,” Journal of Applied prob- ability, vol. 19, no. 1, pp. 221–228, 1982

    work page 1982

  17. [17]

    Concise formulas for the area and volume of a hyper- spherical cap,

    S. Li, “Concise formulas for the area and volume of a hyper- spherical cap,” Asian Journal of Mathematics and Statistics , vol. 4, no. 1, pp. 66–70, 2011

    work page 2011

  18. [18]

    IPFS - Content Addressed, Versioned, P2P File System

    J. Benet, “IPFS-content addressed, versioned, P2P file system,” arXiv preprint arXiv:1407.3561 , 2014

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  19. [19]

    The MD5 message-digest algorithm,

    R. Rivest, “The MD5 message-digest algorithm,” Tech. Rep., 1992

    work page 1992

  20. [20]

    Kec- cak,

    G. Bertoni, J. Daemen, M. Peeters, and G. Van Assche, “Kec- cak,” in Annual international conference on the theory and applications of cryptographic techniques . Springer, 2013, pp. 313–314

    work page 2013

  21. [21]

    Cyclic codes for error detection,

    W. W. Peterson and D. T. Brown, “Cyclic codes for error detection,” Proceedings of the IRE, vol. 49, no. 1, pp. 228–235, 1961

    work page 1961

  22. [22]

    ERC-20 token standard, 2015,

    F. V ogelsteller and V . Buterin, “ERC-20 token standard, 2015,” URL https://github.com/ethereum/EIPs/blob/master/EIPS/eip- 20-token-standard.md, 2018

    work page 2015

  23. [23]

    Learning internal representations by error propagation,

    D. E. Rumelhart, G. E. Hinton, and R. J. Williams, “Learning internal representations by error propagation,” California Univ San Diego La Jolla Inst for Cognitive Science, Tech. Rep., 1985

    work page 1985

  24. [24]

    Filecoin: A decentralized storage network,

    “Filecoin: A decentralized storage network,” URL https:// filecoin.io/filecoin.pdf , 2017

    work page 2017

  25. [25]

    Storj a peer-to-peer cloud storage network,

    S. Wilkinson, T. Boshevski, J. Brandoff, and V . Buterin, “Storj a peer-to-peer cloud storage network,” 2014

    work page 2014

  26. [26]

    Approximate nearest neighbors: towards removing the curse of dimensionality,

    P. Indyk and R. Motwani, “Approximate nearest neighbors: towards removing the curse of dimensionality,” in Proceedings of the thirtieth annual ACM symposium on Theory of computing. ACM, 1998, pp. 604–613

    work page 1998

  27. [27]

    Locality- preserving hashing in multidimensional spaces,

    P. Indyk, R. Motwani, P. Raghavan, and S. Vempala, “Locality- preserving hashing in multidimensional spaces,” in STOC, vol. 97. Citeseer, 1997, pp. 618–625

    work page 1997

  28. [28]

    Locality preserving hashing

    K. Zhao, H. Lu, and J. Mei, “Locality preserving hashing.” in AAAI, 2014, pp. 2874–2881

    work page 2014

  29. [29]

    Similarity estimation techniques from rounding algorithms,

    M. S. Charikar, “Similarity estimation techniques from rounding algorithms,” in Proceedings of the thiry-fourth annual ACM symposium on Theory of computing . ACM, 2002, pp. 380– 388

    work page 2002

  30. [30]

    SingularityNET: A decentralized, open market and inter-network for AIs,

    B. Goertzel, S. Giacomelli, D. Hanson, C. Pennachin, and M. Argentieri, “SingularityNET: A decentralized, open market and inter-network for AIs,” 2017

    work page 2017

  31. [31]

    Effect network: Decen- tralized network for artificial intelligence,

    J. Eisses, L. Verspeek, and C. Dawe, “Effect network: Decen- tralized network for artificial intelligence,” URL http://effect.ai/ download/effect whitepaper.pdf, 2018

    work page 2018

  32. [32]

    Numeraire: A cryptographic token for coordinating machine intelligence and preventing overfitting,

    R. Craib, G. Bradway, X. Dunn, and J. Krug, “Numeraire: A cryptographic token for coordinating machine intelligence and preventing overfitting,” Retrieved, vol. 23, p. 2018, 2017

    work page 2018