Neuromorphic Mimicry Attacks Exploiting Brain-Inspired Computing for Covert Cyber Intrusions

Hemanth Ravipati

REVIEW 5 major objections 5 minor 26 references

Neuromorphic Mimicry Attacks Exploiting Brain-Inspired Computing for Covert Cyber Intrusions

T0 review · 5 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper claims that neuromorphic chips are vulnerable to a new class of covert attacks—Neuromorphic Mimicry Attacks—that exploit probabilistic spike behavior to evade traditional intrusion detection.

desk verdict A plausible-sounding attack concept undermined by a circular simulation, an unspecified baseline detector, and internal inconsistencies; not ready for peer review. read the letter →

arxiv 2505.17094 v1 pith:MB4XXPP2 submitted 2025-05-21 cs.CR cs.LG

Hemanth Ravipati This is my paper · ORCID

classification cs.CRcs.LG

keywords neuromorphiccomputingspikingneuralnetworksmimicryattackssynapticweighttamperinginputpoisoningintrusiondetectionsystemshardwaresecurityadversarialmachinelearning

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 proposes that neuromorphic chips—processors that mimic the brain's spiking neurons, such as Intel's Loihi—carry a new class of security vulnerability: attacks that imitate normal neural activity. The central claim is that by slightly altering synaptic weights or poisoning sensory inputs, an attacker can shift a neuromorphic system's behavior while staying under the radar of traditional intrusion detection. In simulation, synaptic weight tampering succeeded 92% of the time and cut classification accuracy by only 4.8%, and conventional IDS caught just 12% of tampering and 15% of poisoning attacks. The paper argues that this stealth is possible because neuromorphic computation is probabilistic and non-deterministic, so small deviations look like ordinary neural noise. If the claim holds, safety-critical uses like autonomous vehicles, medical implants, and IoT networks would need defenses designed around neural dynamics rather than network traffic.

What carries the argument

The central object is the Neuromorphic Mimicry Attack (NMA), a two-pronged attack on a spiking neural network: tampering with 10% of synaptic weights by ±0.1, or poisoning 5% of sensory inputs with malicious signals. The attack is carried on a 1000-neuron fully connected spiking neural network with random initial weights and spike-timing-dependent plasticity, meant to emulate neuromorphic chips. The stealth mechanism is that the changes stay inside normal variability: accuracy drops by about 4–5%, latency rises 15–25%, and spike-frequency variance moves from 5 to 8 Hz². The proposed countermeasures are neural-specific anomaly detection, which watches spike frequency and weight-change statistics, and secure synaptic learning protocols with cryptographic verification of weight updates.

What would settle it

Run the paper's attack recipes on physical neuromorphic hardware: alter 10% of synaptic weights by ±0.1 and inject 5% malicious sensory inputs during learning, then measure classification accuracy and a standard detector's catch rate. If accuracy drops by much less than 4–5% or a conventional detector catches more than 12–15% of attacks, the quantitative claim does not transfer.

Watch

Extended reading notes

Core claim

The paper's central claim is that Neuromorphic Mimicry Attacks are a viable new class of threat because neuromorphic systems are probabilistic and non-deterministic: an attacker can nudge a small fraction of synapses or inputs and the resulting behavior remains statistically plausible. In the paper's simulation, synaptic weight tampering succeeded in 92% of trials and reduced classification accuracy from 95% to 90.2%, while sensory input poisoning succeeded 87% of the time and reduced accuracy to 90.8%. Traditional intrusion detection caught only 12% of weight-tampering attacks and 15% of input-poisoning attacks, whereas the paper's neural-specific anomaly detector caught 85%. The conclusion the author draws is that security for brain-inspired computing must be built on neural dynamics—spike frequency, synaptic integrity, latency—rather than on conventional network-traffic signatures.

Load-bearing premise

The load-bearing premise is that a 1000-neuron simulated spiking network with random initial weights behaves enough like real neuromorphic hardware that the measured attack success and detection rates will transfer to physical chips.

Editorial extensions

If this is right

Traditional network-based intrusion detection will miss hardware-level neural manipulation, so neuromorphic deployments need monitors that track spike statistics and synaptic integrity.
A 4–5% accuracy drop is enough to mislead an autonomous vehicle or a medical implant while remaining inside normal operating variance, so safety-critical uses need tighter tolerance and redundant validation.
Neural-specific anomaly detection—watching spike-frequency variance and weight-change rates—appears capable of catching most NMAs (85% in simulation) and is a stronger baseline than conventional IDS.
Secure synaptic learning protocols reduce weight-tampering success from 92% to 45%, but input poisoning remains largely effective, pointing to input validation as the next defensive front.
If simulated results transfer to physical neuromorphic hardware, the same recipes should be tested there before deployment in safety-critical edge systems.

Reading between the lines

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

The paper does not run attacks on physical chips; a direct extension would be to repeat the exact tampering and poisoning recipes on available neuromorphic hardware and compare accuracy drop and spike-variance shifts to the simulated values.
Because each attack only lowers accuracy by about 5%, a determined adversary could chain many small misclassifications over time; the single-snapshot metric likely understates cumulative harm in sequential tasks such as navigation.
Input poisoning's success against secure synaptic protocols suggests defenses must also validate the statistical distribution of incoming sensory data, not just protect the weights.
The 85% detection rate still leaves a 15% miss rate; in safety-critical domains, a separate trusted monitor or deterministic sub-system may be needed rather than relying on probabilistic anomaly scores alone.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Colleague,

Bottom line: this one doesn't establish its headline claim. The 'Neuromorphic Mimicry Attack' is presented as a new class of hardware-level threats, but the evaluation is a self-generated Brian2 simulation with no code, data, or hardware validation—and the reported detection rates are computed against a 'traditional IDS' that is never specified. On top of that, the attack numbers contradict themselves: Section V.B reports 92% weight-tampering success, while Figure 3's caption says ~83%.

What's worth keeping: the topic is legitimate. Neuromorphic chips do have a security surface that differs from von Neumann systems, and the two attack vectors—synaptic weight tampering and sensory input poisoning—are plausible entry points. There's also some genuine effort in building a 1000-neuron SNN in Brian2, measuring latency and variance, and sketching countermeasures. If a paper like this were paired with real Loihi or TrueNorth experiments, or even a well-documented simulator with code, the framing could be useful.

The soft spots are severe, though. The central covertness claim is built on an unspecified detector: we never learn what the 'traditional IDS' is, what features it uses, or how the 12–15% detection rate was computed. Given Table I shows the attack and normal classes differ by 10% in spike frequency and 0.7 percentage points in weight change, a fixed-threshold detector would likely do better. More importantly, stealth is imposed by construction: the attacks were crafted to keep accuracy loss under 5%, so the reported 'success' only measures whether the perturbation stayed below that threshold—not whether it evades any real detector. The evaluation is also circular: the same simulated model defines the attacks, generates the dataset, and reports both attack success and detector accuracy. Finally, the reference list is a problem: most citations are to unrelated preprints by a different author (G. Bandarupalli, on topics like Ethereum fraud and SAP migration), which suggests the related-work section was not assembled with genuine engagement.

The paper is not ready for peer review. It might be worth a seminar-style cautionary discussion about how not to validate a security claim, but not as a citable result. If the authors rework it with real hardware or a publicly available simulation, a defined baseline detector, and honest positioning against adversarial ML, I'd revisit it.

Referee Report

5 major / 5 minor

Summary. The manuscript introduces Neuromorphic Mimicry Attacks (NMAs), a proposed class of attacks against neuromorphic hardware, with two attack vectors: synaptic weight tampering and sensory input poisoning. It describes a simulated 1000-neuron spiking neural network in Brian2, a synthetic 10,000-sample dataset, and two countermeasures (neural-specific anomaly detection and secure synaptic learning protocols). The paper reports that weight tampering achieves a 92% success rate, input poisoning an 87% success rate, that traditional intrusion detection systems detect only 12% and 15% of these attacks respectively, and that the proposed anomaly detector reaches 85% detection accuracy.

Significance. If the results were validated on real neuromorphic hardware against a clearly specified baseline detector, the paper would address a genuinely important and understudied area: the security of brain-inspired computing. The conceptual framing of attacks that mimic normal neural activity is worth discussing. However, the current evidence is entirely internal to an uncalibrated simulation, the baseline 'traditional IDS' is undefined, no code or data are provided, and the reported numbers are internally inconsistent. The contribution is therefore prospective rather than demonstrated, and the quantitative claims cannot yet be taken as evidence about real neuromorphic systems.

major comments (5)

[Section V.B, Table I, Table II] The central covertness claim rests on the unsupported statement that 'traditional IDS detecting only 12% of weight tampering and 15% of input poisoning attacks.' No detector algorithm, feature set, threshold, training procedure, or evaluation protocol is specified anywhere in Sections IV or V. This matters because Table I reports statistically significant differences between normal and attack samples (e.g., synaptic weight change 0.5±0.1% vs 1.2±0.3%), and Section IV.A states attacks are generated by altering 10% of weights by ±0.1 or injecting 5% malicious signals. A fixed-threshold detector on the weight-change feature would flag a large fraction of attack samples, so the 12-15% detection figures are not meaningful evidence of covertness without a precise definition of the baseline detector.
[Section IV.B, Section V.A] Stealth is imposed by construction rather than measured. Section IV.B states that NMAs 'reduce system outputs (e.g., classification accuracy) by less than 5% to maintain stealth,' and Section V.A defines attack success rate only as 'percentage of attacks altering classification accuracy.' Consequently, the reported 92% and 87% success rates measure whether the attack stayed below an author-chosen accuracy-degradation threshold, not whether an independent detector would fail to notice the perturbation. The paper does not report detection against an actual detector, so the conclusion in Section V.E that 'NMAs’ stealth' is confirmed is not established.
[Section IV, Figure 3 vs Section V.B] There is a direct internal inconsistency in the central quantitative results. The caption of Figure 3 in Section IV states that input poisoning achieves a slightly higher success rate (~90%) than weight tampering (~83%), while Section V.B reports that weight tampering achieved a 92% success rate and sensory input poisoning an 87% success rate. No explanation is given for the discrepancy, and the figures are not shown in the manuscript. Since these numbers are the paper's headline findings, the inconsistency substantially undermines their reliability.
[Section IV.B] The simulated model is claimed to 'emulate real-world neuromorphic hardware like Intel’s Loihi,' but no calibration, hardware comparison, or transfer argument is provided. The model is a fully connected 1000-neuron Brian2 spiking neural network with random initial weights and STDP learning. There is no evidence that its probabilistic firing behavior, latency characteristics, or weight dynamics match Loihi or TrueNorth. All quantitative results (95% baseline accuracy, drops to 90.2%, spike variance 5 to 8 Hz²) are therefore simulation artifacts unless the model's fidelity is demonstrated.
[Section IV.A, Section V, Section VII.D] The evaluation is entirely internal and non-reproducible. The same simulated model that defines NMAs also generates the attack dataset, and the same model then reports attack success and countermeasure performance. No independent attack or defense is implemented, and no code or data are released despite Section VII.D stating only that they 'will be made on reasonable request.' This circularity, combined with the absence of code, prevents verification of any of the reported rates and makes the results unsuitable for supporting the paper's broad claims about autonomous vehicles, medical implants, and IoT networks.

minor comments (5)

[Section I/References] The reference list contains many citations to the author's own preprints on unrelated topics (e.g., Ethereum fraud detection, legacy system modernization, trade tariffs), while the most relevant neuromorphic security literature is discussed only superficially. The related-work section should instead engage with existing adversarial attacks on spiking neural networks and hardware security evaluations.
[Author affiliation (title page)] The affiliation line states 'Missouri State University, Texas, USA,' which is geographically contradictory; this should be corrected.
[All figures] Figures 1-4 are referenced and captioned in the text, but the actual figures are not included in the manuscript, so the claimed distributions, architecture diagram, and bar graphs cannot be inspected.
[Section IV.B] The baseline classification accuracy of '95% on a simulated IoT sensor dataset' is reported without any description of the task, input encoding, number of classes, or dataset split, so the accuracy figures cannot be interpreted.
[Section V.E] The phrase 'input poisoning is 70% success rate against secure protocols' is grammatically unclear; it should specify whether 70% is the attack success rate or the residual success rate after mitigation.

Circularity Check

1 steps flagged · score 6.0 of 10

Covertness claim is defined into existence: attacks are constrained to reduce accuracy <5%, and that constrained reduction is then reported as evidence of stealth.

self definitional [Section IV.B ('Model Analysis') and Section V.B ('Attack Performance')]
"NMAs were simulated by injecting malicious scripts to alter 10% of synaptic weights by ±0.1 during the learning phase or crafting adversarial inputs to mimic legitimate signals, reducing system outputs (e.g., classification accuracy) by less than 5% to maintain stealth [18, 19]. ... Metrics included attack success rate (percentage of attacks altering classification accuracy) ... Synaptic weight tampering achieved a 92% success rate, reducing classification accuracy by 4.8% (95% to 90.2%)"

The paper's attack-generation protocol imposes the property that is later reported as the main finding. Section IV.B states that NMAs are implemented by perturbations chosen to reduce accuracy by less than 5% 'to maintain stealth,' so any accuracy drop in that range is an input design choice. Section V.A defines attack success as 'percentage of attacks altering classification accuracy,' and Section V.B then reports 4.8% and 4.2% accuracy reductions as evidence of 'high stealth.' The 92%/87% success rates therefore measure only whether the pre-selected small degradation occurred, not whether the perturbation is covert. The conclusion that NMAs evade traditional IDS is not independently established; it is a restatement of the constraint used to generate the attacks.

full rationale

The central quantitative claim of the paper—that NMAs are covert and evade traditional intrusion detection—rests on an attack-construction constraint rather than an independent measurement. The authors explicitly define NMAs as perturbations that reduce classification accuracy by less than 5% to 'maintain stealth,' then report the resulting 4.8%/4.2% accuracy reductions and 92%/87% success rates as evidence of stealth and evasion. This is circular: the reported success rate counts only whether the pre-imposed small degradation occurred, so the 'covert' finding is equivalent to the attack definition. The paper also fails to specify the 'traditional IDS' that detects only 12–15% of attacks, making those numbers unverifiable, though that is a reproducibility gap rather than a demonstrated circular reduction. Because the most load-bearing result reduces by construction, but the paper does contain independent elements such as the proposed countermeasures and the general attack taxonomy, a partial-circularity score of 6 is appropriate rather than a higher score that would require the entire derivation to be definitionally forced.

Assumptions & free parameters 4 free parameters · 4 assumptions · 1 invented entities

The central evaluation depends on the author's simulation choices rather than external benchmarks. Key parameters such as the perturbation magnitudes, poisoning fraction, and stealth threshold are hand-selected, and the dataset itself is generated by the same model used for evaluation. The main invented entity is the NMA concept, which has no independent evidentiary support in the paper.

free parameters (4)

weight tampering magnitude and fraction = +/-0.1 applied to 10% of synaptic weights
Chosen by the authors in Section IV.A/IV.B to make the accuracy drop approximately 4.8% and to keep the attack 'stealthy'.
input poisoning fraction = 5% malicious sensory signals
Chosen in Section IV.A to produce a smaller accuracy drop (4.2%) and maintain covertness.
stealth threshold = less than 5% output accuracy reduction
Section IV.B states NMAs reduce outputs 'by less than 5% to maintain stealth'; this is a hand-picked definition, not an independently measured detection threshold.
normal and attack dataset distribution parameters = normal: 50 +/- 10 Hz, 0.5 +/- 0.1%, 10 +/- 2 ms; attack: 55 +/- 12 Hz, 1.2 +/- 0.3%, 12 +/- 3 ms
Section IV.A reports these as properties of the generated dataset; they are simulation inputs, not measurements from neuromorphic hardware.

assumptions (4)

domain assumption A Brian2 spiking neural network with 1000 fully connected neurons and STDP faithfully emulates real neuromorphic hardware such as Intel Loihi.
Invoked in Section IV.B: 'The model ... emulate real-world neuromorphic hardware like Intel Loihi.' No calibration to hardware is provided.
ad hoc to paper Random sensory inputs to the SNN are representative of IoT, autonomous vehicle, and medical workloads.
Section IV.A says normal activity was generated by feeding 'random sensory inputs' and then generalizes to scenarios such as IoT sensor processing or autonomous vehicle navigation.
domain assumption Traditional intrusion detection systems cannot detect NMAs because they do not analyze spike-level metrics.
Section V.B asserts traditional IDS detect only 12-15% of attacks, but no concrete IDS implementation or dataset is used; this is assumed by design.
standard math Two-sample t-test validity for comparing normal vs attack means.
Section IV.A reports p < 0.05 via t-test; this requires independence and approximate normality, which is not established but is a standard statistical tool.

invented entities (1)

Neuromorphic Mimicry Attack (NMA)
purpose: A named threat class that manipulates synaptic weights and sensory inputs to imitate legitimate neural activity and evade intrusion detection systems.
The paper introduces NMA in Sections I and II but provides no hardware demonstration or falsifiable prediction; all evidence comes from the authors' own simulation.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Neuromorphic Mimicry Attacks Exploiting Brain-Inspired Computing for Covert Cyber Intrusions." pith.science (2026). https://pith.science/paper/MB4XXPP2

@misc{pith2026250517094,
  author       = {Pith},
  title        = {Pith review of: Neuromorphic Mimicry Attacks Exploiting Brain-Inspired Computing for Covert Cyber Intrusions},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MB4XXPP2}},
  note         = {Machine review of arXiv:2505.17094}
}

read the original abstract

Neuromorphic computing, inspired by the human brain's neural architecture, is revolutionizing artificial intelligence and edge computing with its low-power, adaptive, and event-driven designs. However, these unique characteristics introduce novel cybersecurity risks. This paper proposes Neuromorphic Mimicry Attacks (NMAs), a groundbreaking class of threats that exploit the probabilistic and non-deterministic nature of neuromorphic chips to execute covert intrusions. By mimicking legitimate neural activity through techniques such as synaptic weight tampering and sensory input poisoning, NMAs evade traditional intrusion detection systems, posing risks to applications such as autonomous vehicles, smart medical implants, and IoT networks. This research develops a theoretical framework for NMAs, evaluates their impact using a simulated neuromorphic chip dataset, and proposes countermeasures, including neural-specific anomaly detection and secure synaptic learning protocols. The findings underscore the critical need for tailored cybersecurity measures to protect brain-inspired computing, offering a pioneering exploration of this emerging threat landscape.

Figures

Figures reproduced from arXiv: 2505.17094 by the authors.

**Figure 2.** SNN Architecture with Attack Points. Caption: Diagram of the simulated neuromorphic chip, highlighting neurons (circles), synapses (lines), and attack vectors (red arrows) [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Attack Success Rates The bar graph illustrates the success rates of two Neuromorphic Mimicry Attacks. Input poisoning achieves a slightly higher success rate (~90%) compared to weight tampering (~83%). This suggests that injecting malicious sensory inputs is more effective at deceiving neuromorphic systems than altering synaptic weights. Both methods, however, demonstrate high success, emphasizing the vulnerability … view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: Attack Success Rates Table II (from Section IV) shows increased latency (12.5 ms for weight tampering, 11.5 ms for input poisoning) and spike variance under attack, with traditional IDS detecting only 12% of weight tampering and 15% of input poisoning attacks [13]. C. …

**Figure 4.** Figure 4: Latency Impact E. Discussion The findings confirm NMAs’ stealth, with a 92% success rate for weight tampering and only 15% detection by traditional IDS [23]. The dataset’s neural metrics enabled precise analysis, distinguishing this research from network-based datasets…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

26 extracted references · 17 canonical work pages

[1]

Efficient Deep Neural Network for Intrusion Detection Using CIC-IDS-2017 Dataset,

G. Bandarupalli, “Efficient Deep Neural Network for Intrusion Detection Using CIC-IDS-2017 Dataset,” Nov. 2024, doi: 10.21203/RS.3.RS-5424062/V1

doi:10.21203/rs.3.rs-5424062/v1 2017
[2]

Advancing Smart Transportation via AI for Sustainable Traffic Solutions in Saudi 6 Arabia,

G. Bandarupalli, “Advancing Smart Transportation via AI for Sustainable Traffic Solutions in Saudi 6 Arabia,” Nov. 2024, doi: 10.21203/RS.3.RS- 5389235/V1

doi:10.21203/rs.3.rs- 2024
[3]

The Evolution of Blockchain Security and Examining Machine Learning’s Impact on Ethereum Fraud Detection,

G. Bandarupalli, “The Evolution of Blockchain Security and Examining Machine Learning’s Impact on Ethereum Fraud Detection,” Feb. 2025, doi: 10.21203/RS.3.RS-5982424/V1

doi:10.21203/rs.3.rs-5982424/v1 2025
[4]

Enhancing microservices performance with AI-based load balancing: A deep learning perspective,

G. Bandarupalli, “Enhancing microservices performance with AI-based load balancing: A deep learning perspective,” Apr. 2025, doi: 10.21203/rs.3.rs-6396660/v1

work page doi:10.21203/rs.3.rs-6396660/v1 2025
[5]

Enhancing sentiment analysis in multilingual social media data using transformer- based NLP models: A synthetic computational study,

G. Bandarupalli, “Enhancing sentiment analysis in multilingual social media data using transformer- based NLP models: A synthetic computational study,” TechRxiv, Apr. 11, 2025. [Online]. Available: https://doi.org/10.36227/techrxiv.174440282.230131 72/v1

arXiv 2025
[6]

AI-driven code refactoring: Using graph neural networks to enhance software maintainability,

G. Bandarupalli, “AI-driven code refactoring: Using graph neural networks to enhance software maintainability,” arXiv, 2025. [Online]. Available: https://arxiv.org/abs/2504.10412

arXiv 2025
[7]

Code reborn: AI-driven legacy systems modernization from COBOL to Java,

G. Bandarupalli, “Code reborn: AI-driven legacy systems modernization from COBOL to Java,” arXiv,

work page
[8]

SmartSync: Machine Learning for Seamless SAP RAR Data Migration from Legacy ERP Systems,

G. Bandarupalli and V. Kanaparthi, “SmartSync: Machine Learning for Seamless SAP RAR Data Migration from Legacy ERP Systems,” Research Square, Apr. 2025, doi: 10.21203/rs.3.rs-6459008/v1

doi:10.21203/rs.3.rs-6459008/v1 2025

Show all 26 references

[9]

Machine Learning-Driven Analysis of the Economic Impact of Current U.S. Trade Tariffs on Global Supply Chains,

G. Bandarupalli, “Machine Learning-Driven Analysis of the Economic Impact of Current U.S. Trade Tariffs on Global Supply Chains,” TechRxiv, Apr. 17, 2025. [Online]. Available: https://doi.org/10.36227/techrxiv.174490701.174816 32/v1

2025
[10]

Neuromorphic Computing: Opportunities, Challenges, and Applications,

P. K. D. Pramanik et al., “Neuromorphic Computing: Opportunities, Challenges, and Applications,” IEEE Trans. Neural Netw. Learn. Syst., vol. 34, no. 7, pp. 3215–3229, Jul. 2023

2023
[11]

Loihi: A Neuromorphic Manycore Processor with On-Chip Learning,

S. Davies et al., “Loihi: A Neuromorphic Manycore Processor with On-Chip Learning,” IEEE Micro, vol. 38, no. 1, pp. 82–99, Jan. 2018

2018
[12]

Deep Learning in Spiking Neural Networks,

A. Tavanaei et al., “Deep Learning in Spiking Neural Networks,” Neural Netw., vol. 111, pp. 47–63, Mar. 2019

2019
[13]

Security Challenges in Neuromorphic Hardware,

M. Schuman et al., “Security Challenges in Neuromorphic Hardware,” IEEE Secur. Priv., vol. 21, no. 3, pp. 45–53, May 2023

2023
[14]

Toward Generating a New Intrusion Detection Dataset and Intrusion Traffic Characterization,

I. Sharafaldin et al., “Toward Generating a New Intrusion Detection Dataset and Intrusion Traffic Characterization,” in Proc. 4th Int. Conf. Inf. Syst. Secur. Priv., 2018, pp. 108–116

2018
[15]

Intriguing Properties of Neural Networks,

C. Szegedy et al., “Intriguing Properties of Neural Networks,” arXiv, 2013. [Online]. Available: https://arxiv.org/abs/1312.6199

2013 arXiv
[16]

Towards Evaluating the Robustness of Neural Networks,

N. Carlini and D. Wagner, “Towards Evaluating the Robustness of Neural Networks,” in Proc. IEEE Symp. Secur. Priv., 2017, pp. 39–57

2017
[17]

Adversarial Attacks and Defenses in Deep Learning,

J. Chen et al., “Adversarial Attacks and Defenses in Deep Learning,” IEEE Access, vol. 8, pp. 123456– 123467, 2020

2020
[18]

Fairness in Machine Learning: Lessons from Political Philosophy,

R. Binns, “Fairness in Machine Learning: Lessons from Political Philosophy,” J. Mach. Learn. Res., vol. 18, no. 1, pp. 1–11, 2018

2018
[19]

Deep Learning,

Y. LeCun et al., “Deep Learning,” Nature, vol. 521, no. 7553, pp. 436–444, May 2015

2015
[20]

Towards Spike-Based Machine Intelligence with Neuromorphic Computing,

K. Roy et al., “Towards Spike-Based Machine Intelligence with Neuromorphic Computing,” Nature, vol. 575, no. 7784, pp. 607–617, Nov. 2019

2019
[21]

Neuronal Dynamics: From Single Neurons to Networks and Models of Cognition,

W. Gerstner et al., “Neuronal Dynamics: From Single Neurons to Networks and Models of Cognition,” Cambridge Univ. Press, 2014

2014
[22]

Large-Scale Neuromorphic Computing Systems,

S. Furber, “Large-Scale Neuromorphic Computing Systems,” J. Neural Eng., vol. 13, no. 5, 2016

2016
[23]

Advancing Neuromorphic Computing with Loihi,

M. Davies et al., “Advancing Neuromorphic Computing with Loihi,” Intel Technol. J., vol. 22, no. 1, pp. 1–15, 2018

2018
[24]

ImageNet Classification with Deep Convolutional Neural Networks,

A. Krizhevsky et al., “ImageNet Classification with Deep Convolutional Neural Networks,” in Proc. Adv. Neural Inf. Process. Syst., 2012, pp. 1097–1105

2012
[25]

Neuromorphic Silicon Neuron Circuits,

G. Indiveri et al., “Neuromorphic Silicon Neuron Circuits,” Front. Neurosci., vol. 5, no. 73, 2011

2011
[2025]

Available: https://arxiv.org/abs/2504.11335

[Online]. Available: https://arxiv.org/abs/2504.11335

arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Efficient Deep Neural Network for Intrusion Detection Using CIC-IDS-2017 Dataset,

G. Bandarupalli, “Efficient Deep Neural Network for Intrusion Detection Using CIC-IDS-2017 Dataset,” Nov. 2024, doi: 10.21203/RS.3.RS-5424062/V1

doi:10.21203/rs.3.rs-5424062/v1 2017

[2] [2]

Advancing Smart Transportation via AI for Sustainable Traffic Solutions in Saudi 6 Arabia,

G. Bandarupalli, “Advancing Smart Transportation via AI for Sustainable Traffic Solutions in Saudi 6 Arabia,” Nov. 2024, doi: 10.21203/RS.3.RS- 5389235/V1

doi:10.21203/rs.3.rs- 2024

[3] [3]

The Evolution of Blockchain Security and Examining Machine Learning’s Impact on Ethereum Fraud Detection,

G. Bandarupalli, “The Evolution of Blockchain Security and Examining Machine Learning’s Impact on Ethereum Fraud Detection,” Feb. 2025, doi: 10.21203/RS.3.RS-5982424/V1

doi:10.21203/rs.3.rs-5982424/v1 2025

[4] [4]

Enhancing microservices performance with AI-based load balancing: A deep learning perspective,

G. Bandarupalli, “Enhancing microservices performance with AI-based load balancing: A deep learning perspective,” Apr. 2025, doi: 10.21203/rs.3.rs-6396660/v1

work page doi:10.21203/rs.3.rs-6396660/v1 2025

[5] [5]

Enhancing sentiment analysis in multilingual social media data using transformer- based NLP models: A synthetic computational study,

G. Bandarupalli, “Enhancing sentiment analysis in multilingual social media data using transformer- based NLP models: A synthetic computational study,” TechRxiv, Apr. 11, 2025. [Online]. Available: https://doi.org/10.36227/techrxiv.174440282.230131 72/v1

arXiv 2025

[6] [6]

AI-driven code refactoring: Using graph neural networks to enhance software maintainability,

G. Bandarupalli, “AI-driven code refactoring: Using graph neural networks to enhance software maintainability,” arXiv, 2025. [Online]. Available: https://arxiv.org/abs/2504.10412

arXiv 2025

[7] [7]

Code reborn: AI-driven legacy systems modernization from COBOL to Java,

G. Bandarupalli, “Code reborn: AI-driven legacy systems modernization from COBOL to Java,” arXiv,

work page

[8] [8]

SmartSync: Machine Learning for Seamless SAP RAR Data Migration from Legacy ERP Systems,

G. Bandarupalli and V. Kanaparthi, “SmartSync: Machine Learning for Seamless SAP RAR Data Migration from Legacy ERP Systems,” Research Square, Apr. 2025, doi: 10.21203/rs.3.rs-6459008/v1

doi:10.21203/rs.3.rs-6459008/v1 2025