Anomaly Detection and Sampling Cost Control via Hierarchical GANs

Chen Zhong; M. Cenk Gursoy; Senem Velipasalar

REVIEW 3 major objections 4 minor 16 references

Anomaly Detection and Sampling Cost Control via Hierarchical GANs

T0 review · 3 major / 4 minor · reviewed 2026-08-27 · deepseek-v4-flash

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

Pith's one-line read A hierarchy of GANs can predict future values of a stochastic time series, and with a buffer zone around the threshold it detects threshold crossings faster and at lower error cost than a statistics-aware sampling policy, at the price of…

desk verdict The hierarchical GAN idea is plausible, but the main comparison against the known-statistics policy is confounded by an asymmetric buffer-zone alarm rule and no false-alarm metric. read the letter →

arxiv 2009.13598 v1 pith:RB6VPQAH submitted 2020-09-28 cs.LG cs.ITmath.ITstat.ML

Chen Zhong , M. Cenk Gursoy , Senem Velipasalar This is my paper · ORCID

classification cs.LGcs.ITmath.ITstat.ML

keywords anomalydetectionthresholdcrossinghierarchicalGANnonuniformsamplingcostOrnstein-Uhlenbeckprocessbufferzonedelay

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

Threshold crossings in a stochastic time series must be caught by sampling the process, and every sample carries sensing and communication cost. This paper tries to detect crossings without knowing the process statistics: a chain of GANs predicts the next several values from the current sample and uses those predictions to decide when to sample next. A buffer zone around the threshold makes the detector react earlier, and simulations on OU processes show that widening the buffer lowers delay, miss rate, and cost of error at the cost of more samples. The central demonstration is that this learned detector can beat a sampling policy that knows the OU parameters, on delay and sampling ratio, and on error cost once the buffer is wide enough.

What carries the argument

The load-bearing object is the hierarchical GAN predictor: a chain in which GAN $i$ receives the current sample and the predictions of GANs $1$ through $i-1$, $s_j=[x(t),\hat{x}(t+1),\ldots,\hat{x}(t+j-1)]$, and outputs the prediction $\hat{x}(t+j)$; each generator's loss adds the squared prediction error to the adversarial loss. The second load-bearing object is the buffer zone $[\Gamma-\rho,\Gamma+\rho]$: the detector declares a warning on the first predicted sample that falls inside the zone, which makes it fire earlier than waiting for an actual crossing. This buffer is the mechanism that converts a purely sampled process into early detection, and its width $\rho$ is the tuning parameter that controls the tradeoff between sampling cost and detection quality.

What would settle it

Re-run the experiments with the same buffer-zone rule applied to the baseline policy in (3): declare a detection whenever the baseline's sampled value enters $[\Gamma-\rho,\Gamma+\rho]$, and compare delay, miss rate, and cost of error at the same $\rho$ values. If the baseline's metrics improve to match or beat the GAN detector, the claimed advantage is an artifact of the asymmetric alarm rule.

Watch

Extended reading notes

Core claim

The paper's central claim is that a hierarchy of $N$ GANs can act as a nonuniform sampler that predicts the next $N$ values of a stochastic time series from the current sample, and that this predictor, combined with a buffer zone $[\Gamma-\rho,\Gamma+\rho]$ around the threshold, detects threshold crossings with lower delay, miss rate, and cost of error than a sampling policy that knows the process statistics, at the price of sampling more often. In simulations over OU processes with random parameters, widening $\rho$ reduces delay, miss rate, and the cost $\int_{T_{\mathrm{true}}}^{T_{\mathrm{detect}}}|x(t)|\,dt$ in every tested configuration; at $\rho=0.2$ all seven hierarchy lengths tested give lower average error cost than the policy in (3), and most sample less frequently than that policy.

Load-bearing premise

The reported gains depend on comparing the GAN detector, which fires an alarm inside a buffer zone around the target level, against a comparison policy that does not use the same buffer-zone alarm rule; if the comparison policy were allowed the same rule, its delay and miss rates might improve enough to erase the advantage.

Editorial extensions

If this is right

Widening the buffer zone $\rho$ is a reliable way to trade sampling cost for detection quality: in every tested configuration it lowers delay, miss rate, and cost of error while raising the sampling ratio.
With a sufficiently wide buffer, the GAN detector reaches lower average detection delay than the statistics-aware policy in (3) for prediction lengths $N=1,5,10,15,20$.
At $\rho=0.2$, all seven tested hierarchy lengths achieve lower average cost of error than the policy in (3), even though the GAN detector never sees the process parameters.
Prediction length $N$ acts as a second control knob: larger $N$ reduces sampling but increases delay and miss rate, so the two parameters must be chosen jointly for a given operating point.

Reading between the lines

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

A strict control that applies the same buffer-zone alarm rule to the baseline policy in (3) would separate the contribution of GAN predictions from the contribution of the alarm rule; the paper does not report this control.
The observed accumulation of prediction loss with hierarchy level suggests a design rule the paper does not state: add GAN levels only while the marginal loss remains below the delay saving from the longer prediction length.
Because the error cost in (2) weights tail excursions by their size, the buffer width that minimizes that cost will generally not minimize a simple miss-rate metric; applications with other error costs should re-tune $\rho$.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Request a human review

A listed scientist reviews the paper for a fee and the review publishes here regardless of verdict. See the reviewers or get listed.

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

Quick take: the paper has a legitimate new idea—stacking GANs so each level predicts one step ahead and feeds the prediction into the next level—and the internal tradeoff analysis (buffer zone width vs delay/miss/sampling) is internally consistent. But the headline claim, that the GAN detector beats the known-statistics sampling policy in (3) on delay and cost, is not supported by the experiments as designed.

The problem is in Section IV-C. The proposed detector declares a crossing when a predicted sample enters [Γ-ρ, Γ+ρ]. The baseline policy (3) is implemented with no equivalent early-warning rule; it presumably flags only when an actual sample crosses Γ. So a larger ρ makes the GAN detector more permissive, guaranteeing lower measured delay and cost regardless of predictive quality. The paper never applies the same buffer-zone rule to the baseline, and never reports false alarms. Without those, the delay and cost curves in Figs. 4-6 cannot be read as evidence that the GAN's predictions beat the known-statistics policy.

What is genuinely new: the hierarchical GAN structure with level-by-level training and the LSTM-based generator is a reasonable way to extend prediction length without a single GAN blowing up. The observation that lower-level predictions help upper levels is believable, and Fig. 3 shows the error accumulation pattern you'd expect. The buffer-zone parameter ρ as a tunable knob is a sensible engineering idea, even if it is not analyzed theoretically.

The experiments are simple: one synthetic OU process family, parameters drawn from the same range in training and testing, no error bars or significance tests. The miss-rate and cost definitions are explained, but the absence of any false-alarm count is a real gap—a detector that always predicts entry into the buffer zone would look great on delay and misses.

Bottom line: the internal trends (larger ρ reduces delay/miss at increased sampling) are plausible and not confounded. The comparison with policy (3) is the load-bearing claim, and it collapses on the current evidence. The fix is straightforward: give the baseline the same buffer-zone rule, report false alarms, and ideally release code and data.

For peer review: yes, a serious editor should send this out. The idea is worth debating and the flaw is fixable rather than fatal. I just would not let the current comparison stand.

Referee Report

3 major / 4 minor

Summary. This paper proposes a hierarchical GAN-based detector for threshold-crossing anomalies in a stochastic time series with unknown statistics. The detector uses current samples and lower-level GAN predictions to generate future samples, decides when to sample next via Eq. (4), and introduces a buffer zone of width ρ around the threshold to reduce detection delay at the cost of more sampling. Experiments on Ornstein-Uhlenbeck processes compare the proposed detector with the closed-form sampling policy of Eq. (3), which assumes full statistical knowledge. The reported results (Figs. 4–7) claim that the GAN-based detector can achieve lower detection delay, miss rate, and average cost of error for sufficiently large buffer widths, while sampling less frequently than the baseline policy. The paper concludes that the buffer zone provides a tunable tradeoff between detection accuracy and sampling cost.

Significance. If the reported empirical claims are valid, the paper would offer a useful statistics-free alternative to model-based sampling policies for threshold-crossing detection, with a practical tradeoff controlled by the buffer zone. The hierarchical prediction structure and the explicit treatment of sampling cost are sensible ideas. However, the central comparison is currently confounded by an asymmetric alarm rule: the GAN detector declares a crossing when a predicted sample enters the buffer zone, while the baseline policy in Eq. (3) is not given the same early-warning rule. No false-alarm rate is reported, and no error bars or statistical significance tests are provided. The paper contributes no formal proofs or released code; its value is empirical, and the key empirical claim is not yet established in the present form. The stress-test concern raised by the reader is confirmed by the text in Section IV-C and by the experimental design.

major comments (3)

[Section IV-C, Figs. 4–6] The comparison with the policy in Eq. (3) is confounded by an asymmetric alarm rule. The text defines the threshold-crossing time for the proposed detector as the first time instant at which a predicted sample is within [Γ−ρ, Γ+ρ], but the baseline policy in Eq. (3) has no such buffer-zone rule and presumably declares a crossing only when an actual sample crosses Γ. Consequently, the reported reductions in detection delay and miss rate for larger ρ may be direct consequences of the more permissive buffer-zone rule rather than of the GAN's predictive ability. The authors should apply the same buffer-zone alarm rule to the baseline policy, or otherwise make the event definitions identical for both detectors, before claiming superiority in delay and miss rate.
[Section IV-C, Figs. 4–6] No false-alarm metric is reported. A predicted entry into the buffer zone need not be followed by a genuine threshold crossing, so declaring these events as detections can trivially lower the measured delay and average cost of error. The manuscript should report the false-alarm rate or the probability that a detection is not followed by an actual crossing, and it should present the delay/cost performance as a function of, or jointly with, the false-alarm rate.
[Section IV-C, Figs. 4–7] The experimental claims are not supported by any measure of statistical significance. The figures report single curves with no error bars, confidence intervals, or multiple-seed results, yet the text makes quantitative comparisons such as 'three out of the seven tested GAN-based detectors' and 'all seven GAN-based detector can work less costly.' The authors should report means and variances over multiple independent runs, or at least state the number of trials and provide error bars, so that the reported differences can be assessed.

minor comments (4)

[Algorithms 1 and 2] There are typographical errors: 'the the generator' and 'the the discriminator' appear in Algorithm 1 and Algorithm 2, and the email address in the author block has a spacing artifact ('svelipas@sy r.edu').
[Eq. (9)] The squared-error term in the generator loss has no explicitly defined weighting coefficient; the relative importance of the adversarial loss and the prediction loss appears to be fixed at 1, which should be stated and ideally varied in the experiments.
[Section IV-A] The experiment settings omit several implementation details that are important for reproducibility, including the length L of the OU trajectories, the number of training and testing episodes, the LSTM hidden size, the optimizer and learning rate, and the sampling cost c_s used for the baseline policy in Eq. (3).
[Figures 4–7] The figures would be easier to read with distinct markers for each value of N and with a legend, since the curves for different N cross and overlap, especially at small buffer widths.

Circularity Check

1 steps flagged · score 6.0 of 10

Reported delay/miss/cost gains with buffer zone are built into the detection definition itself.

self definitional [Section IV-C, Testing Phase (definition of threshold-crossing time; Figs. 4-6)]
"we introduce a small buffer zone of width ρ around the threshold Γ in the following experiments, and define the threshold-crossing time as the first time instant at which the predicted sample is within the range [Γ −ρ, Γ + ρ]. With this, the delay in detecting a threshold crossing is reduced at the cost of increased number of samples."

The delay metric is defined as Tdetect − Ttrue, and Tdetect is defined as the first time the predicted sample enters [Γ−ρ, Γ+ρ]. For any fixed predicted trajectory, first entry into a wider interval is never later than first entry into a narrower one, so increasing ρ mechanically lowers the measured delay, miss rate, and cost (2), while raising the sampling ratio. This holds independently of the GAN's prediction quality. The baseline policy in (3) is evaluated without any equivalent widened alarm rule, so the reported improvements with ρ and the claimed advantage over policy (3) are partly consequences of the detector definition, not of learned predictions.

full rationale

The hierarchical GAN's core mechanics are not circular: the generators are trained against future samples via the squared-error loss in Eq. (9), and the comparison policy in Eq. (3) comes from independent prior work [15]. There is no fitted parameter renamed as a prediction, no load-bearing self-citation, and no uniqueness claim imported from the authors' own prior results. However, the paper's headline tradeoff — improved detection delay, miss rate, and average cost of error with a larger buffer zone — is produced by the alarm-rule definition itself. Because Tdetect is defined as the first predicted sample in the widened interval, the monotone improvement with ρ is guaranteed by construction for any predictor, and the baseline, lacking this rule, is not compared on equal terms. This makes one central quantitative claim reduce by definition, while the GAN architecture, training loss, and sampling-ratio behavior retain independent content. The score of 6 reflects that partial circularity: one of the paper's main demonstrated results is a definitional consequence rather than an empirical finding about the GAN's predictive capability.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The proposed method introduces no new physical entities. The central claim depends on the OU ground-truth model, the reliability of learned predictions, the validity of the [15] baseline, and the match between training and deployment distributions, none of which are independently verified in the paper.

free parameters (6)

buffer zone width ρ = swept 0 to 0.2
Hand-chosen detection rule parameter; the central delay/sampling tradeoff claim depends on it.
prediction length N (number of GAN levels) = tested at 1, 5, 10, 15, 20, 25, 30
Controls maximum sampling gap; all performance metrics are reported as functions of N.
OU parameter ranges = θ in [0.02,0.03], σ in [0.4,0.6]
Training and test distribution; results may not extend outside this range.
training iterations per GAN = 5000
Training budget chosen without convergence analysis or stopping criterion.
sampling cost c_s in baseline policy = 0.1
Chosen for experiments; determines the baseline sampling frequency.
weight of squared-error term in generator loss (Eq. 9) = 1 (implicit)
Both terms in (9) are added without a tunable coefficient; this choice affects prediction accuracy.

assumptions (5)

domain assumption The monitored process is well modeled by an Ornstein-Uhlenbeck process.
Section II uses Eq. (1) as the ground-truth model for all experiments.
domain assumption A GAN with one LSTM and two fully-connected layers can predict future OU samples from a single current sample and prior predictions.
Section III-B describes the architecture but provides no guarantee or ablations; Fig. 3 shows accumulated errors.
domain assumption The sampling policy in Eq. (3) from [15] is a valid approximately optimal baseline under known statistics.
Section II invokes it as the comparison policy; its approximation quality is not re-derived.
domain assumption Discrete-time simulation with an unspecified time step represents the continuous OU process faithfully.
Section IV uses discrete samples but never states the step size or discretization scheme.
domain assumption Training and test processes drawn from the same parameter ranges are sufficient to demonstrate generalization.
Section IV-A generates both training and test OU processes from the same uniform ranges.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Anomaly Detection and Sampling Cost Control via Hierarchical GANs." pith.science (2026). https://pith.science/paper/RB6VPQAH

@misc{pith2026200913598,
  author       = {Pith},
  title        = {Pith review of: Anomaly Detection and Sampling Cost Control via Hierarchical GANs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RB6VPQAH}},
  note         = {Machine review of arXiv:2009.13598}
}

read the original abstract

Anomaly detection incurs certain sampling and sensing costs and therefore it is of great importance to strike a balance between the detection accuracy and these costs. In this work, we study anomaly detection by considering the detection of threshold crossings in a stochastic time series without the knowledge of its statistics. To reduce the sampling cost in this detection process, we propose the use of hierarchical generative adversarial networks (GANs) to perform nonuniform sampling. In order to improve the detection accuracy and reduce the delay in detection, we introduce a buffer zone in the operation of the proposed GAN-based detector. In the experiments, we analyze the performance of the proposed hierarchical GAN detector considering the metrics of detection delay, miss rates, average cost of error, and sampling ratio. We identify the tradeoffs in the performance as the buffer zone sizes and the number of GAN levels in the hierarchy vary. We also compare the performance with that of a sampling policy that approximately minimizes the sum of average costs of sampling and error given the parameters of the stochastic process. We demonstrate that the proposed GAN-based detector can have significant performance improvements in terms of detection delay and average cost of error with a larger buffer zone but at the cost of increased sampling rates.

Figures

Figures reproduced from arXiv: 2009.13598 by the authors.

**Figure 2.** Structure of hierarchical GAN. B. Hierarchical Structure and Anomaly Detection With the GAN introduced above, the sample x (t) collected at time t can be used to generate the predictions of data in the next N time instants. Therefore, the choice of the value of N determines the maximum gap between the two successive sampling time instants. To control the sampling cost in the anomaly detection, we assume that the sys… view at source ↗

**Figure 3.** The prediction loss in each level of the hierarchical [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Figures from the paper (3 more)

**Figure 5.** Figure 5: The average miss detection ratio vs. buffer zone widt [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]

**Figure 6.** Figure 6: The average cost of error (due to delayed detection) v [PITH_FULL_IMAGE:figures/full_fig_p005_6.png]

**Figure 7.** Figure 7: The average sampling ratio vs. buffer zone width [PITH_FULL_IMAGE:figures/full_fig_p006_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

16 extracted references · 14 canonical work pages

[1]

Anomaly detection in wireless sensor networks,

S. Rajasegarar, C. Leckie, and M. Palaniswami, “Anomaly detection in wireless sensor networks,” IEEE Wireless Communications , vol. 15, no. 4, pp. 34–40, 2008

2008
[2]

Anomaly detection in wirele ss sensor network of the smart home system,

A. Kanev, A. Nasteka, C. Bessonova, D. Nevmerzhitsky, A. Silaev, A. Efremov, and K. Nikiforova, “Anomaly detection in wirele ss sensor network of the smart home system,” in 2017 20th Conference of Open Innovations Association (FRUCT) , pp. 118–124, IEEE, 2017

2017
[3]

Deep reinforcement one-shot le arning for change point detection,

A. Puzanov and K. Cohen, “Deep reinforcement one-shot le arning for change point detection,” in 2018 56th Annual Allerton Conference on Communication, Control, and Computing (Allerton) , pp. 1047–1051, IEEE, 2018

work page 2018
[4]

AD-IoT: Anomaly detection of IoT cyberattacks in smart cit y using machine learning,

I. Alrashdi, A. Alqazzaz, E. Alouﬁ, R. Alharthi, M. Zohdy , and H. Ming, “AD-IoT: Anomaly detection of IoT cyberattacks in smart cit y using machine learning,” in 2019 IEEE 9th Annual Computing and Communi- cation W orkshop and Conference (CCWC) , pp. 0305–0310, 2019

work page 2019
[5]

Ou tlier dirichlet mixture mechanism: Adversarial statistical lea rning for anomaly detection in the fog,

N. Moustafa, K.-K. R. Choo, I. Radwan, and S. Camtepe, “Ou tlier dirichlet mixture mechanism: Adversarial statistical lea rning for anomaly detection in the fog,” IEEE Transactions on Information F orensics and Security, 2019

work page 2019
[6]

Policy d esign for active sequential hypothesis testing using deep learning,

D. Kartik, E. Sabir, U. Mitra, and P . Natarajan, “Policy d esign for active sequential hypothesis testing using deep learning, ” in 2018 56th Annual Allerton Conference on Communication, Control, and Computing (Allerton), pp. 741–748, IEEE, 2018

work page 2018
[7]

Deep actor-c ritic re- inforcement learning for anomaly detection,

C. Zhong, M. C. Gursoy, and S. V elipasalar, “Deep actor-c ritic re- inforcement learning for anomaly detection,” in 2019 IEEE Global Communications Conference (GLOBECOM) , pp. 1–6, 2019

work page 2019
[8]

Time Series Anomaly Detection Using Convolutional Neural Networks and Transfer Learning

T. Wen and R. Keyes, “Time series anomaly detection using convolutional neural networks and transfer learning,” arXiv preprint arXiv:1905.13628, 2019

work page Pith review arXiv 1905

Show all 16 references

[9]

A deep neural network for unsupervised anomaly detection and diagnosis in multivari ate time series data,

C. Zhang, D. Song, Y . Chen, X. Feng, C. Lumezanu, W. Cheng, J. Ni, B. Zong, H. Chen, and N. V . Chawla, “A deep neural network for unsupervised anomaly detection and diagnosis in multivari ate time series data,” in Proceedings of the AAAI Conference on Artiﬁcial Intelligen ce...

2019
[10]

Time series anomaly detection based on GAN,

Y . Sun, W. Y u, Y . Chen, and A. Kadam, “Time series anomaly detection based on GAN,” in 2019 Sixth International Conference on Social Networks Analysis, Management and Security (SNAMS) , pp. 375–382, 2019

2019
[11]

MAD-G AN: Multi- variate anomaly detection for time series data with generat ive adversarial networks,

D. Li, D. Chen, B. Jin, L. Shi, J. Goh, and S.-K. Ng, “MAD-G AN: Multi- variate anomaly detection for time series data with generat ive adversarial networks,” in Artiﬁcial Neural Networks and Machine Learning – ICANN 2019: Text and Time Series (I. V . Tetko, V . K˚ urkov´ a, P...

2019
[12]

Active ano maly detection with switching cost,

D. Chen, Q. Huang, H. Feng, Q. Zhao, and B. Hu, “Active ano maly detection with switching cost,” in ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICA SSP), pp. 5346–5350, IEEE, 2019

2019
[13]

Sequential anomaly detection under a nonlinear system cost,

A. Gurevich, K. Cohen, and Q. Zhao, “Sequential anomaly detection under a nonlinear system cost,” IEEE Transactions on Signal Processing , vol. 67, no. 14, pp. 3689–3703, 2019

2019
[14]

Signiﬁc ant sampling for shortest path routing: A deep reinforcement learning so lution,

Y . Shao, A. Rezaee, S. C. Liew, and V . W. S. Chan, “Signiﬁc ant sampling for shortest path routing: A deep reinforcement learning so lution,” in 2019 IEEE Global Communications Conference (GLOBECOM) , pp. 1– 7, 2019

2019
[15]

Cognitive network manageme nt and control with signiﬁcantly reduced state sensing,

A. Rezaee and V . W. S. Chan, “Cognitive network manageme nt and control with signiﬁcantly reduced state sensing,” IEEE Transactions on Cognitive Communications and Networking , vol. 5, no. 3, pp. 783–794, 2019

2019
[16]

Generative adversar ial nets,

I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. War de-Farley, S. Ozair, A. Courville, and Y . Bengio, “Generative adversar ial nets,” in Advances in Neural Information Processing Systems 27 (Z. Ghahramani, M. Welling, C. Cortes, N. D. Lawrence, and K. Q. Weinberger, e ds...

2014

Pith tools

Reviewed August 27, 2026 · model on record in the stance chip above.

[1] [1]

Anomaly detection in wireless sensor networks,

S. Rajasegarar, C. Leckie, and M. Palaniswami, “Anomaly detection in wireless sensor networks,” IEEE Wireless Communications , vol. 15, no. 4, pp. 34–40, 2008

2008

[2] [2]

Anomaly detection in wirele ss sensor network of the smart home system,

A. Kanev, A. Nasteka, C. Bessonova, D. Nevmerzhitsky, A. Silaev, A. Efremov, and K. Nikiforova, “Anomaly detection in wirele ss sensor network of the smart home system,” in 2017 20th Conference of Open Innovations Association (FRUCT) , pp. 118–124, IEEE, 2017

2017

[3] [3]

Deep reinforcement one-shot le arning for change point detection,

A. Puzanov and K. Cohen, “Deep reinforcement one-shot le arning for change point detection,” in 2018 56th Annual Allerton Conference on Communication, Control, and Computing (Allerton) , pp. 1047–1051, IEEE, 2018

work page 2018

[4] [4]

AD-IoT: Anomaly detection of IoT cyberattacks in smart cit y using machine learning,

I. Alrashdi, A. Alqazzaz, E. Alouﬁ, R. Alharthi, M. Zohdy , and H. Ming, “AD-IoT: Anomaly detection of IoT cyberattacks in smart cit y using machine learning,” in 2019 IEEE 9th Annual Computing and Communi- cation W orkshop and Conference (CCWC) , pp. 0305–0310, 2019

work page 2019

[5] [5]

Ou tlier dirichlet mixture mechanism: Adversarial statistical lea rning for anomaly detection in the fog,

N. Moustafa, K.-K. R. Choo, I. Radwan, and S. Camtepe, “Ou tlier dirichlet mixture mechanism: Adversarial statistical lea rning for anomaly detection in the fog,” IEEE Transactions on Information F orensics and Security, 2019

work page 2019

[6] [6]

Policy d esign for active sequential hypothesis testing using deep learning,

D. Kartik, E. Sabir, U. Mitra, and P . Natarajan, “Policy d esign for active sequential hypothesis testing using deep learning, ” in 2018 56th Annual Allerton Conference on Communication, Control, and Computing (Allerton), pp. 741–748, IEEE, 2018

work page 2018

[7] [7]

Deep actor-c ritic re- inforcement learning for anomaly detection,

C. Zhong, M. C. Gursoy, and S. V elipasalar, “Deep actor-c ritic re- inforcement learning for anomaly detection,” in 2019 IEEE Global Communications Conference (GLOBECOM) , pp. 1–6, 2019

work page 2019

[8] [8]

Time Series Anomaly Detection Using Convolutional Neural Networks and Transfer Learning

T. Wen and R. Keyes, “Time series anomaly detection using convolutional neural networks and transfer learning,” arXiv preprint arXiv:1905.13628, 2019

work page Pith review arXiv 1905