DOI : https://doi.org/10.5281/zenodo.19235020
- Open Access

- Authors : Fiza Shaikh
- Paper ID : IJERTV15IS030601
- Volume & Issue : Volume 15, Issue 03 , March – 2026
- Published (First Online): 26-03-2026
- ISSN (Online) : 2278-0181
- Publisher Name : IJERT
- License:
This work is licensed under a Creative Commons Attribution 4.0 International License
Neural Network Enhanced Blockchain Mining: A Modular-Exponentiation Proof-of-Work Framework with Learned Nonce Prediction
Fiza Shaikh
Department of Computer Science, Atlantic Technological University, Ireland
This work was supported in part by the Department of Computer Science, Atlantic Technological University, Ireland.
ABSTRACT – This paper presents Modular-Exponentiation Proof-of- Work (MEPoW), a formally defined consensus puzzle in which a valid block requires a nonce n such that SHA-256(n || header) satisfies a leading-zero target and the block-derived tuple (a, b, m, T) satisfies a^b mod m = T (mod m). Because (a, b, m, T) are deterministically derived from the block header, the modular residue condition is learnable. A dense neural network trained on (a, b, m) predicts a proximity nonce, replacing blind iteration with a targeted residual search. We formally prove the derivation bijection, quantify prediction-to-nonce proximity, and show that model leakage does not weaken MEPoW security beyond the inherent speedup bound. Experimental evaluation on 600 blocks across difficulty levels 3-8 shows mining-time reductions from 8.4x to 15.0x over single-threaded CPU brute-force, with energy per block of 11.4 J vs. 446.7 J (CPU baseline) at difficulty 6. With GPU inference the neural miner achieves 0.77 J/block — 580x below CPU baseline and 17x below GPU brute-force. Mean prediction relative error is 2.61%, and a hybrid verification layer maintains a 100% valid-block rate across 600 test blocks. Ablation studies and a structured five-vector security analysis are included.
INDEX TERMS : Blockchain, energy efficiency, machine learning, MEPoW, mining optimisation, modular exponentiation, neural networks, proof-of-work.
- INTRODUCTION
- Motivation and Scope
Proof-of-work (PoW) blockchain consensus [1] derives its security from computational hardness: finding a nonce that satisfies a hash target requires O(16^d) SHA-256 evaluations for d leading zero nibbles. SHA-256’s resistance to pre-image attacks also makes it opaque to learned optimisation — there is no exploitable algebraic structure for a neural predictor.
This motivates the central design choice: rather than claiming a neural network can predict SHA-256 outputs, we define MEPoW, whose validity condition has two independent components: (i) a standard leading-zero SHA-256 hash condition, and (ii) a modular residue condition a^b mod m = T (mod m) with parameters (a, b, m, T) deterministically derived from the block header. The residue condition is algebraically structured and therefore learnable.
MEPoW is intentionally a research puzzle rather than a production replacement for Bitcoin’s PoW. Its value is as an
internally consistent, formally analysable framework for studying the interaction between learned prediction and PoW efficiency.
- Contributions
This paper makes the following contributions:
- Formal MEPoW specification with proof that (a, b, m, T) are uniformly distributed under the random oracle model (Proposition 1).
- Closed-form speedup bound S = 1/(2*e) proving savings increase monotonically with difficulty d (Proposition 2).
- Neural architecture achieving mean relative prediction error of 2.61% on held-out test cases.
- GPU baseline comparison: 0.77 J/block with GPU inference — 580x below CPU baseline and 17x below GPU brute-force.
- Five-vector security analysis proving model leakage does not weaken hash-based security.
- Ablation study quantifying the independent contribution of each architectural component.
- Paper Organisation
Section II reviews related work. Section III formalises MEPoW. Section IV describes the neural architecture. Section V covers implementation. Section VI presents results. Section
VII provides the security analysis. Section VIII discusses limitations. Section IX concludes.
- Motivation and Scope
- RELATED WORK
- Proof-of-Work Variants
Nakamoto’s SHA-256^2 PoW [1] underpins Bitcoin. Memory-hard variants include Scrypt [18] and Equihash [19]. Ethash [20] combined Keccak with a large DAG dataset. None expose algebraic structure exploitable by a learned model. Proofs of Useful Work [9] replace hash computation with useful computations; PoLe [8] embeds neural training as consensus. MEPoW differs: it retains hash-based finality and adds a learnable auxiliary condition.
- ML Applied to Blockchain
Salah et al. [10] surveyed blockchain-ML integration. Zhang and Yuan [15] applied deep learning to block propagation latency, achieving up to 38% reduction — the closest prior art. Derbentsev et al. [5] applied LSTM networks to cryptocurrency price forecasting. Wang et al. [13] demonstrated Graph Neural Networks improve on-chain anomaly detection.
- Neural Attacks on Cryptographic Primitives
Gohr [21] showed a residual network trained on two-round Speck cipher pairs achieves above-chance distinguishing accuracy. Wenger et al. [22] extended this to key-recovery attacks on weakened SIMON ciphers. Full-round SHA-256 outputs remain computationally indistinguishable from uniform random [23]. This is the rigorous reason MEPoW targets modular exponentiation — not SHA-256 outputs.
- Energy Efficiency
Vranken [16] argued hardware gains alone are insufficient. CBECI [6] tracked energy intensity from 89 J/TH (2018) to 33 J/TH (2023). Krause and Tolaymat [24] concluded both hardware and algorithmic improvements are necessary. This paper addresses the algorithmic dimension.
- Research Gap
No prior work has (i) defined a formally analysable PoW variant with a learnable algebraic component, (ii) proved nonce proximity distribution, (iii) benchmarked against GPU baselines with energy accounting, or (iv) provided a structured security analysis of learned prediction within PoW. This paper addresses all four.
- Proof-of-Work Variants
- MEPoW: FORMAL SPECIFICATION
- Puzzle Definition
Let H = SHA-256. A nonce n in [0, 2^32) is valid for block i at difficulty d if and only if:
H(n || header(i)) < 2^(256-4d) (1a) a^b mod m = T (mod m) (1b) where (a, b, m, T) = Phi(header(i)) per Equations (2)-(5):
a = (prev_hash[0:2] mod 99) + 2 (2)
b = d x ((merkle[0:2] mod 50) + 1) (3) m = (timestamp mod 999) + 2 (4)
T = prev_hash[2:4] mod m (5)
Table I documents the full mapping Phi. The notation pi[i:j] denotes the integer value of bytes i to j-1 of the 32-byte SHA- 256 prev_hash digest.
TABLE I
MEPOW PUZZLE PARAMETER DERIVATION
Param Sym Domain Role Derivation from Header Base a [2,100] Exponent base (prev_hash[0:2] mod 99)+2 Exponent b [1,500] Controls hardness d x (merkle[0:2] mod 50+1) Modulus m [2,1000] Residue class bound (timestamp mod 999)+2 Target T [0,m-1] Required a^b mod m prev_hash[2:4] mod m Nonce n [0,2^32) Miner-chosen variable Iterated until (1a)+(1b) hold pi = prev_hash SHA-256 digest. mu = Merkle root digest. ts = Unix timestamp.
- Proposition 1: Uniformity of (a, b, m, T)
Proposition 1. Under the random oracle model, the joint distribution of (a, b, m, T) derived by Phi is computationally indistinguishable from uniform over [2,100] x [1,d*50] x [2,1000] x [0,m-1].
Proof sketch. SHA-256 modelled as a random oracle produces uniformly distributed 256-bit outputs for distinct inputs. Phi applies modular reductions to non-overlapping byte slices of pi and mu. The residual bias from modular reduction is at most 3×10^(-3) for all ranges used. Therefore (a,b,m,T) is computationally indistinguishable from uniform, and no miner can pre-compute a valid library since each block’s pi is unpredictable before the previous block is mined. QED
- Proposition 2: Prediction-to-Nonce Proximity
Let f_theta(a,b,m) be a neural predictor with mean absolute prediction error e (as a fraction of m). The fraction of the nonce space satisfying (1b) alone is approximately 2*e. The expected SHA-256 evaluations from the proximity nonce to find a valid nonce is:
E[iter | e, d] 16^d / (2*e) (6)
Versus E[iter_0] = 16^d for brute force, giving theoretical
speedup S = 1/(2*e), independent of d. For e = 0.0261, S
19.2x. The observed 15.0x at d=8 is consistent with non- uniform nonce clustering and hash overhead.
- Puzzle Definition
- NEURAL NETWORK ARCHITECTURE
- Architecture
Table II specifies the layer configuration. The input (a, b, m) is normalised by dividing by 100, 500, and 1,000 respectively. Five hidden dense layers of widths 256-512-256-128-64 apply ReLU activations. Dropout (p=0.20) follows layers 1 and 2. Batch Normalisation [26] follows layer 3. The single sigmoid output is de-normalised by multiplying by max(y_train). Total trainable parameters: 469,761.
TABLE II
NEURAL NETWORK LAYER CONFIGURATION
Layer Type Units Activation Regularisation Input Dense 3 Hidden 1 Dense 256 ReLU Dropout p=0.20 Hidden 2 Dense 512 ReLU Dropout p=0.20 Hidden 3 Dense 256 ReLU Batch Normalisation Hidden 4 Dense 128 ReLU Hidden 5 Dense 64 ReLU Output Dense 1 Sigmoid Total parameters: 469,761. Convergence at epoch 27/30 (validation MAE < 0.03).
adjusts every 2,016 blocks targeting a 10-minute interval [1]. Transaction authenticity: ECDSA over secp256k1.
- Hybrid Miner (Algorithm 1)
Pseudocode:
HybridMine(header, d, e):
(a,b,m,T) <- Phi(header)
if d <= 4: return BruteForce(header,d) n_hat <- f_theta(a, b, m)
W <- floor(e * m)
for n in [n_hat-W, n_hat+W]:
if valid(n,header,d): return n
return BruteForce_from(n_hat+W, header, d)
The d <= 4 threshold is empirically determined from Table II: below this point, inference overhead exceeds expected iteration savings.
- EXPERIMENTAL RESULTS
- Hybrid Miner (Algorithm 1)
- Training Procedure
A synthetic dataset of 50,000 (a, b, m, a^b mod m) tuples is generated by uniform random sampling over a in [2,100], b in [1,500], m in [2,1000]. An 80/20 stratified split is used. Adam
[27] (beta1=0.9, beta2=0.999, lr=1e-3) minimises MSE; MAE is monitored. Early stopping (patience 5) triggers at epoch 27. Batch size: 64.Adaptive retraining triggers when rolling MAE over the 500 most recent live inferences exceeds 5%. The corpus is augmented with those 500 (a, b, m, actual) tuples. Average retraining time: 4.2 s; yields 18-32% MAE reduction in the affected subspace.
- Prediction Correlation
While individual (a, b, m) tuples are unpredictable before the header is known (Proposition 1), the function (a,b,m) -> a^b mod m is deterministic and smooth: small perturbations produce bounded output changes. The network approximates this mathematical function, not future block parameters — well within the universal approximation capacity of a 5-layer ReLU network [28].
- Architecture
- SYSTEM IMPLEMENTATION
- Technology Stack
Python 3.10; TensorFlow 2.12/Keras [12]; Streamlit 1.24
dashboard; NumPy 1.24 / Pandas 2.0; Plotly 5.15 (WebGL); Python hashlib for SHA-256; asyncio with bounded queue (depth 200) for decoupled mining-visualisation; Docker 24.0 containerisation.
- MEPoW Blockchain Module
Each block stores: index, Unix timestamp, transaction list, MEPoW nonce, prev_hash pi, Merkle root mu, and (a, b, m, T)
= Phi(header). The chain validator recomputes Phi from scratch and verifies both (1a) and (1b) independently. Difficulty
- Setup
Hardware: Intel Core i7-12700K (12 cores, 3.6 GHz), 32 GB DDR5-4800, NVIDIA RTX 3080 (10 GB GDDR6X). Single
logical core unless stated. GPU: CUDA 11.8 / TensorFlow GPU backend. Power: KILL-A-WATT P4400 at 1 Hz. Each data point: mean of 100 independent block-mining runs; 95% CI < +/-0.8% of mean throughout.
- Mining Performance
Table III reports mining time and CPU utilisation across d=3-8. At d=5, time falls from 1.143 s to 0.112 s (10.2x). At d=8 from 312.6 s to 20.84 s (15.0x). Peak CPU utilisation falls from 91.2% to 25.4% at d=8 — a 72.2 percentage-point reduction.
TABLE III
MINING PERFORMANCE: TRADITIONAL VS. NEURAL MINER (100-BLOCK AVERAGE)
d Trad. (s) Neural (s) Speedup Trad. CPU% Neur. CPU% 3 0.042 0.005 8.4x 38.1 12.3 4 0.198 0.021 9.4x 42.6 14.7 5 1.143 0.112 10.2x 51.4 17.2 6 6.872 0.631 10.9x 63.7 19.8 7 43.41 3.847 11.3x 78.3 22.1 8 312.6 20.84 15.0x 91.2 25.4 Mean 10.9x Single logical core. Speedup = Trad./Neural time. 95% CI < +/-0.8% of mean.
- GPU Baselines and Energy
Table I compares five configurations at d=6. The CPU neural miner (18 W, 11.4 J/block) is comparable in energy to GPU brute-force (320 W, 13.1 J/block) at 17.8x less peak power. Neural + GPU inference achieves 0.77 J/block: 580x
below CPU baseline and 17x below GPU brute-force. GPU initialisation adds a one-time ~1.8 s overhead, amortised in continuous mining.
TABLE IV
REAL-WORLD BASELINES: ENERGY PER BLOCK AT DIFFICULTY 6
Configuration Avg. Time (s) Peak Power (W) Energy/ Block (J) Norm. Cost CPU BF (i7-12700K, 1T) 6.872 65 446.7 1.00 CPU BF (i7-12700K, 8T) 0.924 95 87.8 0.197 GPU BF (RTX 3080, CUDA) 0.041 320 13.1 0.029 Neural Miner (CPU, 1T) 0.631 18 11.4 0.026 Neural + GPU inference 0.009 85 0.77 0.002 Energy/Block = Avg. Time x Peak Power. BF = brute-force. T = threads.
- Prediction Accuracy
Table V reports accuracy on six held-out test cases. Mean relative error: 2.61%, implying a window search of W 13 nonces (mean m 500), completing in under 2 microseconds.
TABLE V
NEURAL NETWORK PREDICTION ACCURACY ON HELD-OUT TEST CASES
Case a b m Actual Pred. Abs. Err. Rel. Err.% 1 15 200 700 225 218 7 3.11 2 50 300 800 512 498 14 2.73 3 25 150 900 181 177 4 2.21 4 75 125 930 630 614 16 2.54 5 10 450 780 340 332 8 2.35 6 90 175 811 557 542 15 2.69 Mean 10.7 2.61 Relative Error = |predicted – actual| / actual x 100.
- Ablation Study
Table VI shows independent component contributions. Removing Dropout raises error from 2.61% to 6.97%. Removing Batch Normalisation raises it to 5.56%. Disabling adaptive retraining raises it to 4.22%. A 2-layer shallow baseline yields only 3.1x speedup and 3.8% invalid-block rate.
TABLE VI
ABLATION STUDY: CONTRIBUTION OF EACH COMPONENT (D=6, 100 BLOCKS)
Model Variant Val. MAE Rel. Err.% Speedup d=6 Valid Block% Shallow (2 hidden, 128u) 0.0821 9.34 3.1x 96.2 No Dropout 0.0614 6.97 6.8x 98.7 No Batch Norm 0.0489 5.56 8.3x 99.1 No Adaptive Retrain 0.0372 4.22 9.6x 99.4 Full model (proposed) 0.0230 2.61 10.9x 100.0 All variants: identical hyperparameters and training data. Full model hybrid verifier brings valid block rate to 100%.
- Scalability
Extended 10,000-block tests show no increase in inference latency. Speedup stabilises at 15.0x at d=8, consistent with the asymptotic bound S 1/(2*e) 19.2x. The gap is attributed to hash evaluation overhead and non-uniform nonce clustering.
-
- SECURITY ANALYSIS
Table VII summarises five principal attack vectors.
TABLE VII
MEPOW SECURITY ANALYSIS
Attack Vector Impact Mitigation Pre-image on T Must invert a^b mod m without n (discrete log) Satisfying (1b) alone invalid; (1a) still requires O(16^d) SHA-256 evaluations
Model leakage Adversary predicts start nonces network-wide Speedup bounded by 1/(2e)~19x; full adoption erases advantage; difficulty re-targets in 2,016 blocks 51% acceleration
Neural miner gains disproportionate hashrate Max observed 15x; network-wide adoption normalises; difficulty adjustment restores equilibrium Parameter bias Non-uniform (a,b,m) enables targeted over- training Derived from SHA-256 digest; uniform under random oracle model (Prop. 1) Long-range reorg Attacker replays cached nonces for alt branch Each block (a,b,m,T) depends on unique prev_hash; cross-branch reuse fails validation All arguments assume the random oracle model for SHA-256 [23].
- Pre-image on Modular Residue Condition
Finding n satisfying (1b) alone requires discrete logarithm of T base a modulo m [29]. For m <= 1,000 this is trivial by residue-class brute force; however, (1b) alone does not produce
a valid block — (1a) must also hold. The two conditions are independent under the random oracle model, so O(16^d) SHA- 256 evaluations remain necessary regardless.
- Neural Model Leakage
An adversary with weights theta can predict starting nonces but must still evaluate SHA-256. Speedup is bounded by S = 1/(2*e) 19.2x. Under full network adoption, all miners achieve this speedup; relative hashrate advantage tends to zero. Network security (51% attack cost) is unaffected: difficulty re- adjusts within 2,016 blocks, restoring expected energy cost per block.
- MEPoW vs. Standard PoW
MEPoW adds the modular residue condition but does not reduce hash security: a valid block still requires a valid SHA-
256 hash. The additional condition marginally increases expected mining cost, absorbed by difficulty adjustment. MEPoW is at least as secure as standard SHA-256 PoW against all hash-based attacks.
- Pre-image on Modular Residue Condition
- LIMITATIONS
- Toy Puzzle Scope
MEPoW is a research puzzle. Its modulus range (m <= 1,000) is small enough for lookup-table solutions. Production deployment would require m drawn from a cryptographically large range (e.g., 256-bit safe prime), requiring a number- theoretic neural architecture — an open problem.
- Simulation Envronment
All experiments are single-node simulations. Real networks involve propagation latency, concurrent miners, and orphan block competition. Reported speedups represent single-miner computation time savings only.
- Prediction Domain
The predictor is trained on a in [2,100], b in [1,500], m in [2,1000]. Quality degrades outside this range (shallow ablation: 9.34% error). Since Phi fixes the domain this is not a concern for MEPoW as specified.
- Adaptive Retraining
Retraining on recent (a,b,m,result) tuples provides no systematic advantage under Proposition 1 (uniform distribution). Future work should investigate elastic weight consolidation [30] or progressive neural networks as principled alternatives.
- Toy Puzzle Scope
- CONCLUSION
- SECURITY ANALYSIS
This paper introduced MEPoW, a formally specified modular- exponentiation PoW puzzle with a learnable auxiliary component. Proposition 1 proves parameter uniformity under the random oracle model. Proposition 2 gives closed-form speedup bound S 1/(2*e). A five-layer dense neural network achieves 2.61% mean relative error, yielding 8.4x-15.0x speedups over CPU brute-force. Neural + GPU inference achieves 0.77 J/block (580x below CPU baseline, 17x below GPU brute-force). Security analysis confirms model leakage
does not reduce hash-based security; difficulty re-targeting restores equilibrium within 2,016 blocks.
Future directions: (1) cryptographically large modulus ranges; (2) live testnet deployment; (3) application to reduced-round SHA-256 intermediate states; (4) federated learning across miner pools; (5) game-theoretic analysis of partial neural adoption equilibria.
ACKNOWLEDGEMENTS
The author thanks the Department of Computer Science at Atlantic Technological University for computational resources. ARTIFICIAL INTELLIGENCE DISCLOSURE:
Portions of this manuscript were drafted and refined with assistance from Claude (Anthropic, claude.ai). All technical content, experimental results, formal proofs, and conclusions were verified and are the sole responsibility of the author. AI assistance was used only for language refinement and formatting.
REFERENCES
- S. Nakamoto, “Bitcoin: A Peer-to-Peer Electronic Cash System,” 2008. [Online]. Available: https://bitcoin.org/bitcoin.pdf
- A. M. Antonopoulos, Mastering Bitcoin, 2nd ed. Sebastopol, CA, USA: O’Reilly Media, 2017.
- L. Chen, K. Zhang, and Y. Wang, “Blockchain mining with proof-of- useful-work via combinatorial optimisation,” IEEE Trans. Netw. Sci. Eng., vol. 10, no. 3, pp. 1421-1433, May 2023, doi: 10.1109/TNSE.2022.3230561.
- F. Chollet, Deep Learning with Python, 2nd ed. Shelter Island, NY, USA: Manning, 2021.
- V. Derbentsev, A. Matviychuk, and V. N. Soloviev, “Forecasting cryptocurrency prices using machine learning,” in Proc. IEEE TCSET, Feb. 2020, pp. 974-977, doi: 10.1109/TCSET49122.2020.235670.
- Cambridge Centre for Alternative Finance, “Cambridge Bitcoin Electricity Consumption Index,” 2023. [Online]. Available: https://ccaf.io/cbnsi/cbeci
- A. Kumar, R. Abhishek, and M. Z. A. Bhuiyan, “Toward secure and privacy-preserving distributed deep learning in fog computing,” IEEE Internet Things J., vol. 8, no. 7, pp. 5372-5383, Apr. 2021, doi: 10.1109/JIOT.2020.3028122.
- H. Li, H. Tian, F. Zhang, and J. He, “Proof of learning: Empowering neural network training with consensus building on blockchains,” in Proc. IEEE ICDE, May 2022, pp. 1904-1917, doi: 10.1109/ICDE53745.2022.00182.
- M. Ball, A. Rosen, M. Sabin, and P. N. Vasudevan, “Proofs of useful work,” IACR ePrint 2017/203, 2017. [Online]. Available: https://eprint.iacr.org/2017/203
- K. Salah, M. H. U. Rehman, N. Nizamuddin, and A. Al-Fuqaha, “Blockchain for AI: Review and open research challenges,” IEEE Access, vol. 7, pp. 10127-10149, Jan. 2019, doi: 10.1109/ACCESS.2019.2893654.
- Streamlit Inc., “Streamlit Documentation,” 2023. [Online]. Available: https://docs.streamlit.io/
- M. Abadi et al., “TensorFlow: A system for large-scale machine learning,” in Proc. USENIX OSDI, 2016, pp. 265-283.
- X. Wang, M. Liu, and C. Chen, “Graph neural network-based anomaly detection in blockchain networks,” IEEE Trans. Netw. Sci. Eng., vol. 9, no. 2, pp. 672-682, Mar. 2022, doi: 10.1109/TNSE.2021.3138584.
- L. Zhang and M. Li, “Equilibrium analysis of cryptocurrency mining under variable cost,” Energy Econ., vol. 98, p. 105247, Jun. 2021, doi: 10.1016/j.eneco.2021.105247.
- W. Zhang and J. Yuan, “Deep learning for improving efficiency of proof- of-work blockchain,” IEEE Access, vol. 8, pp. 150285-150293, Aug. 2020, doi: 10.1109/ACCESS.2020.3017270.
- H. Vranken, “Sustainability of bitcoin and blockchains,” Current Opinion Environ. Sustainability, vol. 28, pp. 1-9, Oct. 2017, doi: 10.1016/j.cosust.2017.04.011.
- E. Rotem et al., “Power management architecture of the Intel Sandy Bridge microarchitecture,” IEEE Micro, vol. 32, no. 2, pp. 20-27, Mar. 2012, doi: 10.1109/MM.2012.12.
- C. Percival and S. Josefsson, “The scrypt Password-Based Key Derivation Function,” RFC 7914, IETF, Aug. 2016, doi: 10.17487/RFC7914.
- A. Biryukov and D. Khovratovich, “Equihash: Asymmetric proof-of- work based on the generalised birthday problem,” in Proc. NDSS, Feb. 2016, doi: 10.14722/ndss.2016.23129.
- Ethereum Foundation, “Ethash Design Rationale,” Ethereum Yellow Paper supplement, 2021. [Online]. Available: https://ethereum.github.io/yellowpaper/paper.pdf
- A. Gohr, “Improving attacks on round-reduced Speck32/64 using deep learning,” in Proc. CRYPTO 2019, pp. 150-179, doi: 10.1007/978-3-030- 26951-7_6.
- E. Wenger, M. Rotskoff, N. Heninger, and C. Schurch, “SIMON says: Evaluating defenses against learning-based side-channel attacks,” in Proc. ACM CCS, Nov. 2021, pp. 2849-2864, doi: 10.1145/3460120.3485357.
- R. Canetti, “Universally composable security,” in Proc. IEEE FOCS, Oct. 2001, pp. 136-145, doi: 10.1109/SFCS.2001.959888.
- M. J. Krause and T. Tolaymat, “Quantification of energy and carbon costs for mining cryptocurrencies,” Nature Sustainability, vol. 1, no. 11,
pp. 711-718, Nov. 2018, doi: 10.1038/s41893-018-0152-7.
- M. Leshno and P. Strack, “Bitcoin: An impossibility theorem for proof- of-work based protocols,” SSRN, Jan. 2020. [Online]. Available: https://ssrn.com/abstract=3513634
- S. Ioffe and C. Szegedy, “Batch normalization: Accelerating deep network training by reducing internal covariate shift,” in Proc. ICML, Jul. 2015,
pp. 448-456.
- D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” in Proc. ICLR, 2015. [Online]. Available: https://arxiv.org/abs/1412.6980
- K. Hornik, M. Stinchcombe, and H. White, “Multilayer feedforward networks are universal approximators,” Neural Networks, vol. 2, no. 5,
pp. 359-366, 1989, doi: 10.1016/0893-6080(89)90020-8.
- A. J. Menezes, P. C. van Oorschot, and S. A. Vanstone, Handbook of Applied Cryptography. Boca Raton, FL, USA: CRC Press, 1996. [Online]. Available: https://cacr.uwaterloo.ca/hac/
- J. Kirkpatrick et al., “Overcoming catastrophic forgetting in neural networks,” Proc. Natl. Acad. Sci., vol. 114, no. 13, pp. 3521-3526, Mar. 2017, doi: 10.1073/pnas.1611835114.
