Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Specification
The disclosure is objected to because of the following informalities:
Paragraph 0032 refers to FIG. 1 as FIG. 2 and it should be FIG.1 as it explains the reference numbers in FIG. 1.
Appropriate correction is required.
Claim Rejections - 35 USC § 101
35 U.S.C. 101 reads as follows:
Whoever invents or discovers any new and useful process, machine, manufacture, or composition of matter, or any new and useful improvement thereof, may obtain a patent therefor, subject to the conditions and requirements of this title.
Claims 1-20 rejected under 35 U.S.C. 101 because the claimed invention is directed to abstract idea without significantly more. The claim(s) recite(s) significantly more. The subject matter eligibility test for products and process is describe below for claim 1 in view of dependent claims.
Regarding claim 1:
Step 1: Is the claim to a process machine manufacture or composition of matter?
Yes – Claim 1 recites a method, which is a method that falls under the statutory categories.
Step 2A Prong 1: Does the claim recite an abstract idea, law of nature, or natural phenomenon?
Yes – The claim recites the following:
“and comparing the reconstructed time series with the input time series to determine whether the input time series is out-of-distribution with the in-distribution time series.” - The limitations of recites the metal process of comparing if reconstruction was in or out of distribution (see MPEP 2106.04(a)(2)III).
Step 2 Prong 2: Does the claim recite additional elements that integrate the judicial exception into a particular application? No –
The claim includes the additional element(s):
“A method comprising steps of: receiving an input time series;”
The additional elements fall under Insignificant Extra-Solution Activity as mere data gathering by receiving in input time series. See MPEP 2106.5(g).
“causing random imputations in the input time series to provide an imputed time series;”
The additional elements fall under “apply it” as using a generic computer to cause random imputations in the time series. See Mere Instructions to Apply an Exemption (see MPEP 2106.05(f)).
“processing the imputed time series with a diffusion model that has been parameterized on a given in-distribution time series to obtain a reconstructed time series;”
The additional elements fall under “apply it” as using a generic computer to process the imputed time series with a diffusion model and obtain a reconstructed time series (see MPEP 2106.05(f)).
Step 2B: Does the claim recite additional elements that amount to significantly more than the judicial exception?
No - The claim does not include additional elements that are sufficient to amount to a significantly more than the judicial exemption. As an order whole, the claim is directed to the use of diffusion models to reconstruct time series data. As discussed above with respect to integration of the abstract idea into a practical application, the additional elements of receiving causing, and processing fall under using generic computer to apply an exemption and mere data gathering. The method does not improve on the function of a computer, transforms an article into another article, nor is it applied by a particular machine, making the claim not patent eligible.
Regarding claim 2:
Step 2A Prong 1: Does the claim recite an abstract idea, law of nature, or natural phenomenon?
Yes – The claim recites the following:
“wherein the comparing includes determining a score based on distance between the reconstructed time series and the input time series,” - The limitation recites a mathematical process of calculating a score by determining a distance between the reconstructed time series and input time series (see MPEP 2106.04(a)(2)I)
Step 2A Prong 2, Step 2B: The additional element(s):
“wherein the score is an indicator of a likelihood the input time series is out-of-distribution.”
The additional elements fall under Insignificant Extra-Solution Activity. See MPEP 2106.5(g). The judicial exemptions do not integrate into a practical application nor provide an improvement. The process does not provide an inventive concept nor provides a practical application.
Regarding claim 3:
Step 2A Prong 1: Does the claim recite an abstract idea, law of nature, or natural phenomenon?
Yes – The claim recites the following:
“The method of claim 2, wherein the distance is one of Euclidean distance and cosine similarity.” - The limitation recites a mathematical process Euclidean distance and cosine similarity (see MPEP 2106.04(a)(2)I)
Step 2A Prong 2, Step 2B: The additional element(s):
No additional elements. The judicial exemptions do not integrate into a practical application nor provide an improvement. The process does not provide an inventive concept nor provides a practical application.
Regarding claim 4:
Step 2A Prong 2, Step 2B: The additional element(s):
“The method of claim 1, wherein, when the input time series is out-of-distribution, the reconstructed time series is a bad reconstruction relative to when the input time series is in-distribution.”
The additional elements fall under Insignificant Extra-Solution Activity. See MPEP 2106.5(g). The judicial exemptions do not integrate into a practical application nor provide an improvement. The process does not provide an inventive concept nor provides a practical application.
Regarding claim 5:
Step 2A Prong 2, Step 2B: The additional element(s):
“The method of claim 1, wherein the processing further includes: utilizing domain-specific side information with the imputed time series in the diffusion model.”
The additional elements fall under “apply it” as using a generic computer to utilize domain-specific side information to imputed the time series (see MPEP 2106.05(f)). The judicial exemptions do not integrate into a practical application nor provide an improvement. The process does not provide an inventive concept nor provides a practical application.
Regarding claim 6:
Step 2A Prong 2, Step 2B: The additional element(s):
“The method of claim 1, wherein the random imputations are performed using a mask determined based on a number of time steps and a number of features.”
The additional elements fall under “apply it” as using a generic computer to us a mask to determine a number of features (see MPEP 2106.05(f)). The judicial exemptions do not integrate into a practical application nor provide an improvement. The process does not provide an inventive concept nor provides a practical application.
Regarding claim 7:
Step 2A Prong 2, Step 2B: The additional element(s):
“The method of claim 1, wherein the steps further include: training the diffusion model with in-distribution time series data, such that the comparing determines whether or not the input time series belongs to a same distribution as the in-distribution time series data”
The additional element falls under the “apply it” by using computers to train the diffusion model (MPEP 2106.05(f)). The judicial exemptions do not integrate into a practical application nor provide an improvement. The process does not provide an inventive concept nor provides a practical application.
Regarding claim 8:
Step 2A Prong 1: Does the claim recite an abstract idea, law of nature, or natural phenomenon?
Yes – The claim recites the following:
“The method of claim 1, wherein the steps further include: classifying the input time series based on the comparing such that (1) when the input time series is in-distribution, the input time series is classified as belonging to a domain associated with the in-distribution time series, and (2) when the input time series is out-of-distribution, the input time series is classified as not belonging to the domain.” - The limitations recites metal process when a time series is consider in-distribution and when it is out-distribution will be classifying as associated with a domain (see MPEP 2106.04(a)(2)III).
Step 2A Prong 2, Step 2B: The additional element(s):
No additional elements. See MPEP 2106.5(g). The judicial exemptions do not integrate into a practical application nor provide an improvement. The process does not provide an inventive concept nor provides a practical application.
Regarding claim 9:
Step 2A Prong 2, Step 2B: The additional element(s):
“The method of claim 1, wherein the steps further include: determining whether the input time series is anomalous Internet of Things (loT) communications based on the comparing.”
The additional element falls under the “apply it” by using computers to determine when the input time series is anomalous IoT based on comparing (MPEP 2106.05(f)). The judicial exemptions do not integrate into a practical application nor provide an improvement. The process does not provide an inventive concept nor provides a practical application.
Regarding claim 10:
Step 2A Prong 2, Step 2B: The additional element(s):
“The method of claim 1, wherein the steps further include: determining whether the input time series corresponds to a Distributed Denial of Service (DDoS) network flow based on the comparing.”
The additional element falls under the “apply it” by using computers to determine when the input time series is anomalous DDoS based on comparing (MPEP 2106.05(f)). The judicial exemptions do not integrate into a practical application nor provide an improvement. The process does not provide an inventive concept nor provides a practical application.
Claims 11-20 recite a computer readable medium product and are analogous to the method of claims 1-10. Therefore, the rejections of claim 1-10 above applies to claims 11-20.
Claim Rejections - 35 USC § 103
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
Claim(s) 1, 2, 4, 6, 7, 8, 11, 12, 14, 16, 17 and 18 are rejected under 35 U.S.C. 103 as being unpatentable over Chen, Yuhang, et al. "Imdiffusion: Imputed diffusion models for multivariate time series anomaly detection." arXiv preprint arXiv:2307.00754 (2023 (“Chen”) in view of Graham, Mark S., et al. "Denoising diffusion models for out-of-distribution detection." Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2023. (“Graham”).
Regarding claim 1 and analogous claim 11, Chen teaches A method comprising steps of: receiving an input time series; causing random imputations in the input time series to provide an imputed time series (Chen page 3 3 PRELIMINARY In this section, we present the problem of MTS anomaly detection and time series imputation, and an overview of diffusion models.
3.1 Multivariate Time Series Anomaly Detection para 1 line 1-3, We consider a collection of MTS denoted as X, which encompasses measurements recorded from timestamp 1 to 𝐿. Specifically:
PNG
media_image1.png
31
367
media_image1.png
Greyscale
[receiving an input time series;]
Page 3 3.3 Denoising Diffusion Model, Our ImDiffusion is based on the diffusion models [63], a well known generative model that draws inspiration from non-equilibrium thermodynamics. Diffusion models follow a two-step process for data generation. Firstly, it introduces noise to the input incrementally, akin to a forward process. Secondly, it learns to generate new samples by progressively removing the noise from a sample noise vector, thereby resembling a reverse process… Here, 𝛽 is a positive constant that can either be learned or predefined, representing the noise level. It is important to note that the forward process is parameterized as a Markov chain, as the values of X𝑡 solely depend on X𝑡−1. The final step, X𝑇 , is fully corrupted and becomes random noise [causing random imputations in the input time series].
Page 4 4.1 Imputed Diffusion Models paragraph 5 line 10-14, The lower subplot in Figure 2 demonstrates the application of an unconditional diffusion model. A notable distinction from the conditional model (upper subplot) is the substantial difference in imputed error between normal and abnormal data points.
Page 5,
PNG
media_image2.png
305
563
media_image2.png
Greyscale
[to provide an imputed time series]);
processing the imputed time series with a diffusion model [that has been parameterized on a given in-distribution] time series to obtain a reconstructed time series (Chen page 2 Figure 1,
PNG
media_image3.png
560
759
media_image3.png
Greyscale
Page 2 1. Introduction para 4 line 9-19, The imputation-based approach employed by ImDiffusion offers distinct advantages over forecasting and reconstruction methods. Firstly, it leverages neighboring values in the time series as additional conditional information, enabling a more accurate modeling of the temporal and inter-correlated dependencies present in multivariate data. Secondly, the reference information from neighboring values helps to reduce uncertainty in predictions, thereby enhancing the robustness of the detection process. Figure 1 presents an example in which forecasting, reconstruction, and imputation methods are employed to predict a time series using diffusion models [processing the imputed time series with a diffusion model].
Page 3 Chen page 3 para 2 line 1-4, Conversely, we employ a machine learning model with learnable parameters Θ to denoise X𝑇 and reconstruct X0. This is accomplished by iteratively computing the following Gaussian transitions:
PNG
media_image4.png
56
685
media_image4.png
Greyscale
[obtain a reconstructed time series]);
Chen does not explicitly teach [processing the imputed time series with a diffusion model] that has been parameterized on a given in-distribution [time series to obtain a reconstructed time series]; and comparing the reconstructed time series with the input time series to determine whether the input time series is out-of-distribution with the in-distribution time series.
However Graham teaches [processing the imputed time series with a diffusion model] that has been parameterized on a given in-distribution [time series to obtain a reconstructed time series] (Graham page 2950, 3. Method para 1 line 1-13,
To enable reconstruction-based OOD detection that is not dependent on a fixed information bottleneck, we propose making use of a trained DDPM [10] to reconstruct images. During training, samples x0 are degraded according to a fixed process with Gaussian noise according to a timestep t and a noise variance schedule βt to produce noised samples xt, such that
PNG
media_image5.png
53
478
media_image5.png
Greyscale
where Q
Q
≤
t
≤
T
and we define αt := 1 − βt and
α
¯
≔
∏
s
=
1
t
α
s
. The schedule βt is designed to increase with t and have the property that the fully noised xT is close to an isotropic Gaussian, xT ∼ N (0, I); i.e. xT contains no information about.
page 2950, 3. Method para 2,
Given this trained model and a test input x0, we can sample a set of xt for a range of values of t and estimate their reconstructions,
x
^
0
,
t
=
p
θ
(
x
0
|
x
t
)
. Measuring the similarity between each reconstruction and input
S
(
x
^
0
,
t
,
x
t
)
provides a range of similarity scores that can be used to decide whether x0 is in-distribution.
Page 2951 4.1 Experimental details, Datasets. We evaluate our method both on a number of common computer vision benchmarks and, recognizing that performance on these benchmarks don’t necessarily reflect performance in the real world, on a set of higher dimension medical imaging datasets. For the computer vision benchmarks we use four in-distribution datasets: Fashion- MNIST [41], CIFAR10 [16], CelebA [19], and SVHN [26]. For the grayscale FashionMNIST, we use MNIST [17] as an OOD dataset. For the other colour datasets, we use all other colour datasets as OOD datasets. CelebA images were resized to 32x32 to match the dimension of the other colour datasets.We also use vertically- and horizontally-flipped versions of each in-distribution dataset as further OOD datasets, giving a total of 15 pairs of in vs out-of-distribution datasets [that has been parameterized on a given in-distribution]);
and comparing the reconstructed time series with the input time series to determine whether the input time series is out-of-distribution with the in-distribution time series (Graham page 2948,
PNG
media_image6.png
691
594
media_image6.png
Greyscale
[and comparing the reconstructed time series with the input time series to]
page 2950- 2951, 3.3. Evaluating similarity It is common to use the mean-squared error (MSE) between the input and reconstruction to evaluate similarity. In this work, we also choose to use the LPIPS metric, which uses the distance between the deep features of a network (in this case, Alexnet [15]) from two inputs as a measure of their perceptual similarity. LPIPS has been shown to correlate well with human evaluations of image similarity [47]. Using both MSE and LPIPS gives a total of 2N similarity measurements per input for the N reconstructions performed. We convert each measurement into a Z-score using the measurements from a validation set for each reconstruction and metric (MSE or LPIPS) separately. We average these 2N Z-scores to produce an OOD score for each input.
Implementation details. For our method, we used the DDPM model as described in [31]1. We used a 3-layer UNet with channels, with two residual blocks per [256, 512, 784] layer and a single-headed attention block after each residual block in layers 2,3 of the downsampling branch and layer 1 of the upsampling branch. The timestep was sinusoidally embedded and passed through a two-layer MLP with a Swish activation function [29] to create a 1024-dim embedding. We used during training and a linear noise schedule T = 1000 with varying between 0.0015 and 0.0195. All models ßt were trained for 300 epochs using the Adam optimiser [12] with a learning rate of 2.5e-5. At test time, we used the PLMS sampler with 100 timesteps and reconstructed from each of these 100 steps as starting points to produce 100 reconstructions per input.
Page 2950, 4.2. Results for computer vision datasets Results are presented in Table 1, reported as AUC scores. The DDPM has the highest average rank across all experiments, with the state-of-the-art DoSE ranked second-highest. In a direct comparison, the DDPM outperforms DoSE on 13/15 dataset pairings. The increases in performance afforded by the DDPM are sometimes substantial, most notably when the OOD dataset is a vertically- or horizontally flipped version of the in-distribution dataset.
Page 2955 4.6. Variants of the proposed method We investigated the effects of changing the method used to classify OOD samples. We first tried using a one-class SVM or a Gaussian Mixture Model (GMM) to score samples. Results in Table 3 show that, while all methods perform well, the GMM and SVM are consistently outperformed by the simple Z-score averaging method. [determine whether the input time series is out-of-distribution with the in-distribution time series]).
Chen and Graham are considered to be analogous to the claim invention because they are in the same field of diffusion models. Therefore, it would have been obvious to someone of ordinary skill in the art before the effective filling date of the claimed invention to have modified Chen to incorporate the teachings of Graham to train the model with in-distribution samples to detect out-of-distribution samples. Doing so to classify out-of-distribution samples in and unsupervised method (Graham page 2948 Abstract line 15-22, We propose to use DDPMs to reconstruct an input that has been noised to a range of noise levels, and use the resulting multi-dimensional reconstruction error to classify out-of-distribution inputs. We validate our approach both on standard computer-vision datasets and on higher dimension medical datasets. Our approach outperforms not only reconstruction-based methods, but also state of-the-art generative-based approaches.
Page 2955 5. Conclusion line 1-6, In this work, we explored how DDPMs can be used to perform unsupervised OOD detection. We propose performing reconstructions of a number of inputs noised to different extents, addressing a drawback of standard reconstruction techniques that require the choice of a single, fixed bottleneck.).
Regarding claim 2 and analogous claim 12, Chen in view of Graham teach the method of claim 1 and analogous claim 11.
Chen and Graham are combine in the same rational as set forth above with respect to claim 1 and analogous claim 11.
Graham teaches wherein the comparing includes determining a score based on distance between the reconstructed time series and the input time series, wherein the score is an indicator of a likelihood the input time series is out-of-distribution (Graham page 2950- 2951, 3.3. Evaluating similarity It is common to use the mean-squared error (MSE) between the input and reconstruction to evaluate similarity. In this work, we also choose to use the LPIPS metric, which uses the distance between the deep features of a network (in this case, Alexnet [15]) from two inputs as a measure of their perceptual similarity. LPIPS has been shown to correlate well with human evaluations of image similarity [47]. Using both MSE and LPIPS gives a total of 2N similarity measurements per input for the N reconstructions performed. We convert each measurement into a Z-score using the measurements from a validation set for each reconstruction and metric (MSE or LPIPS) separately. We average these 2N Z-scores to produce an OOD score for each input [wherein the comparing includes determining a score based on distance between the reconstructed time series and the input time series] (i.e. the z-score indicates how likely that a particular example is OOD)).
Regarding claim 4 and analogous claim 14, Chen in view of Graham teach the method of claim 1 and analogous claim 11.
Chen and Graham are combine in the same rational as set forth above with respect to claim 1 and analogous claim 11.
Graham teaches wherein, when the input time series is out-of-distribution, the reconstructed time series is a bad reconstruction relative to when the input time series is in-distribution (Graham page 2952, 4.2. Results for computer vision datasets para 2, Fig. 2 shows some reconstructions from the model trained on the SVHN dataset. Reconstructions from all four models are included in Supplementary. Reconstructions of the in-distribution SVHN input still retain similarity to the input up until noising to t = 500 − 600, whilst the OOD reconstructions start to look dissimilar to their inputs after noising to t = 100 and bear almost no resemblance by t = 400 [when the input time series is out-of-distribution]. The plot also shows that for t ⪆ 700 the noised images retain very little information from the input and the model outputs begin to resemble unconditioned samples more than they do reconstructions. This suggests that reconstructions from higher t contribute little to the OOD signal and can potentially be discarded, though we view it as an advantage of our method that it performs well across dataset pairings without any post-hoc need for selecting the range of values to reconstruct from. We explore this further in Sec. 4.5.
2953 Figure 2
PNG
media_image7.png
460
1144
media_image7.png
Greyscale
[the reconstructed time series is a bad reconstruction relative to when the input time series is in-distribution.] (i.e. the reconstruction for 2-5 rows are OOD and have bad reconstruction compared to when the sample is in-distribution)).
Regarding claim 6 and analogous claim 16, Chen in view of Graham teach the method of claim 1 and analogous claim 11.
Chen further teaches wherein the random imputations are performed using a mask determined based on a number of time steps and a number of features (Chen Page 3,
PNG
media_image8.png
264
471
media_image8.png
Greyscale
Page 5, 4.2 Design of Data Masking
The ImDiffusion approach leverages deliberate masking, using a maskMapplied to the time series data, to create unobserved points that require imputation. The choice of the masking strategy plays a crucial role in determining the performance of anomaly detection. In this paper, we compare two masking strategies: Random strategy: This strategy randomly masks data values in the raw time series with a 50% probability [34]. It provides a straightforward and simple masking technique [wherein the random imputations are performed using a mask].
Page 6 algorithm 1 line 3-7,
PNG
media_image9.png
129
445
media_image9.png
Greyscale
[mask determined based on a number of time steps and a number of features]).
Regarding claim 7 and analogous claim 17, Chen in view of Graham teach the method of claim 1 and analogous claim 11.
Chen and Graham are combine in the same rational as set forth above with respect to claim 1 and analogous claim 11.
Graham further teaches wherein the steps further include: training the diffusion model with in-distribution time series data, such that the comparing determines whether or not the input time series belongs to a same distribution as the in-distribution time series data (Graham page 2950- 2951, 3.3. Evaluating similarity It is common to use the mean-squared error (MSE) between the input and reconstruction to evaluate similarity. In this work, we also choose to use the LPIPS metric, which uses the distance between the deep features of a network (in this case, Alexnet [15]) from two inputs as a measure of their perceptual similarity. LPIPS has been shown to correlate well with human evaluations of image similarity [47]. Using both MSE and LPIPS gives a total of 2N similarity measurements per input for the N reconstructions performed. We convert each measurement into a Z-score using the measurements from a validation set for each reconstruction and metric (MSE or LPIPS) separately. We average these 2N Z-scores to produce an OOD score for each input.
Page 6 Figure 2.
PNG
media_image10.png
344
886
media_image10.png
Greyscale
[training the diffusion model with in-distribution time series data,].
Implementation details.
For our method, we used the DDPM model as described in [31]1. We used a 3-layer UNet with channels, with two residual blocks per [256, 512, 784] layer and a single-headed attention block after each residual block in layers 2,3 of the downsampling branch and layer 1 of the upsampling branch. The timestep was sinusoidally embedded and passed through a two-layer MLP with a Swish activation function [29] to create a 1024-dim embedding. We used during training and a linear noise schedule T = 1000 with varying between 0.0015 and 0.0195. All models ßt were trained for 300 epochs using the Adam optimiser [12] with a learning rate of 2.5e-5. At test time, we used the PLMS sampler with 100 timesteps and reconstructed from each of these 100 steps as starting points to produce 100 reconstructions per input.
Page 2950, 4.2. Results for computer vision datasets Results are presented in Table 1, reported as AUC scores. The DDPM has the highest average rank across all experiments, with the state-of-the-art DoSE ranked second-highest. In a direct comparison, the DDPM outperforms DoSE on 13/15 dataset pairings. The increases in performance afforded by the DDPM are sometimes substantial, most notably when the OOD dataset is a vertically- or horizontally flipped version of the in-distribution dataset.
4.6. Variants of the proposed method
We investigated the effects of changing the method used to classify OOD samples. We first tried using a one-class SVM or a Gaussian Mixture Model (GMM) to score samples. Results in Table 3 show that, while all methods perform well, the GMM and SVM are consistently outperformed by the simple Z-score averaging method [such that the comparing determines whether or not the input time series belongs to a same distribution as the in-distribution time series data]).
Regarding claim 8 and analogous claim 18, Chen in view of Graham teach the method of claim 1 and analogous claim 11.
Chen and Graham are combine in the same rational as set forth above with respect to claim 1 and analogous claim 11.
Graham teaches wherein the steps further include: classifying the input time series based on the comparing such that (1) when the input time series is in-distribution, the input time series is classified as belonging to a domain associated with the in-distribution time series, and (2) when the input time series is out-of-distribution, the input time series is classified as not belonging to the domain (Graham page 2948,
PNG
media_image11.png
517
440
media_image11.png
Greyscale
[classifying the input time series based on the comparing such that (1) when the input time series is in-distribution, the input time series is classified as belonging to a domain associated with the in-distribution time series,]
Page 2955, Table 3. AUC score for variants of the DDPM method. In the top set of results, the similarity metrics used are fixed to MSE+LPIPS, and we try three classification methods: a GMM, SVM, and a Z-score average. In the bottom set, the classification method is fixed to Z-score averaging, and we use three different similarity metrics: MSE, LPIPS, and MSE+LPIPS.
Page 2955 4.6. Variants of the proposed method para 1 line1-7
We investigated the effects of changing the method used to classify OOD samples. We first tried using a one-class SVM or a Gaussian Mixture Model (GMM) to score samples. Results in Table 3 show that, while all methods perform well, the GMM and SVM are consistently outperformed by the simple Z-score averaging method [and (2) when the input time series is out-of-distribution, the input time series is classified as not belonging to the domain.]).
Claim(s) 3 and 13 are rejected under 35 U.S.C. 103 as being unpatentable over Chen in view of Graham and further in view of Nitsch, Julia, et al. "Out-of-distribution detection for automotive perception." 2021 IEEE International Intelligent Transportation Systems Conference (ITSC). IEEE, 2021 (“Nitsch”).
Regarding claim 3 and analogous claim 13, Chen in view of Graham teach the method of claim 2 and analogous 12.
Chen and Graham are combine in the same rational as set forth above with respect to claim 1 and analogous claim 11.
Chen does not explicitly teach wherein the distance is one of Euclidean distance and cosine similarity.
However Nitsch teaches wherein the distance is one of Euclidean distance and cosine similarity (Nitsch page 2941 B. Post Hoc Network Statistics line Having trained the network para 1 line 1-12, we now want to compute the parameters of class-conditioned Gaussian distributions over the logits, also referred to as the bottleneck vector, of the classification network. The class-conditioned Gaussian is a reasonable modeling choice for the parameter distribution according to Lee et al. [24]. The distance to the Gaussian distribution deemed most likely by the softmax distribution is used to identify OOD samples. We first consider the cosine similarity measure, which is a known distance metric for comparing high dimensional feature vectors and, thus, an effective means for OOD detection:
PNG
media_image12.png
47
412
media_image12.png
Greyscale
C. Quantitative Results line1-12
Table I and Table II show the results on the automotive datasets, KITTI (Din) and nuScenes (Din), with the ImageNet test set serving as Dout. The best result is highlighted in bold and the second best in italics. The results in Table I and Table II confirm that the joint loss scheme is superior to the regular NN training for OOD detection. Furthermore, the cosine similarity is amongst the best results in all experiments with a clear advantage on the AUPRinmetric. When considering the post hoc approaches to OOD detection on their own (CE term in Eq. (3)), we found the cosine similarity to still be the superior choice for standard CE loss trained classifiers [cosine similarity].)).
Chen and Nitsch are considered to be analogous to the claim invention because they are in the same field of machine learning. Therefore, it would have been obvious to someone of ordinary skill in the art before the effective filling date of the claimed invention to have modified Chen to incorporate the teachings of Nitsch to use cosine similarity. Doing so to provide superior out-of-distribution detection using cosine similarity (Nitsch Page 2941 C. Quantitative Results line1-12 Table I and Table II show the results on the automotive datasets, KITTI (Din) and nuScenes (Din), with the ImageNet test set serving as Dout. The best result is highlighted in bold and the second best in italics. The results in Table I and Table II confirm that the joint loss scheme is superior to the regular NN training for OOD detection. Furthermore, the cosine similarity is amongst the best results in all experiments with a clear advantage on the AUPRinmetric. When considering the post hoc approaches to OOD detection on their own (CE term in Eq. (3)), we found the cosine similarity to still be the superior choice for standard CE loss trained classifiers).
Claim(s) 5 and 15 are rejected under 35 U.S.C. 103 as being unpatentable over Chen in view of Graham and further in view of Levac, Brett, et al. "MRI reconstruction with side information using diffusion models." 2023 57th Asilomar Conference on Signals, Systems, and Computers. IEEE, 2023. (“Levac”).
Regarding claim 5 and analogous claim 15, Chen in view of Graham teach the method of claim 1 and analogous claim 11.
Chen and Graham are combine in the same rational as set forth above with respect to claim 1 and analogous claim 11.
Chen does not explicitly teach wherein the processing further includes: utilizing domain-specific side information with the imputed time series in the diffusion model.
However Levac teaches wherein the processing further includes: utilizing domain-specific side information with the imputed time series in the diffusion model (Levac Page 5, IV. RESULTS
Numerical results for reconstructing x2 in the joint reconstruction case are shown in Table I for each approach described above. Example reconstructions for R = 3 are shown in Fig. 5. The top and bottom images in the left most column are xGT2 and xGT1 respectively. The remaining columns each represent a different reconstruction technique as described in Section III-C. For each reconstruction technique the top row is the reconstructed image and the bottom row contains the difference image between the reconstructed image and the ground truth image. All images are plotted on the same dynamic range and the difference images are scaled by a factor of 10.
Page 6 Fig. 5,
PNG
media_image13.png
340
925
media_image13.png
Greyscale
).
Chen and Levac are considered to be analogous to the claim invention because they are in the same field of machine learning. Therefore, it would have been obvious to someone of ordinary skill in the art before the effective filling date of the claimed invention to have modified Chen to incorporate the teachings of Levac to include side information. Doing so to use Bayesian prior and score-based generative model to provide improvement in data reconstruction fidelity (Levac Abstract line 14-20, this work we show that by learning a joint Bayesian prior over multi-contrast data with a score-based generative model we are able to leverage the underlying structure between random variables related to a given imaging problem. This leads to an improvement in image reconstruction fidelity over generative models that rely only on a marginal prior over the image contrast of interest.).
Claim(s) 9, 10, 19 and 20 are rejected under 35 U.S.C. 103 as being unpatentable over Chen in view of Graham and further in view of Pourhmadi et al. (W02022/243980A1) (“Pourhmadi”).
Regarding claim 9 and analogous claim 19, Chen in view of Graham teach the method of claim 1 and analogous claim 11.
Che and Graham are combine in the same rational as set forth above with respect to claim 1 and analogous claim 11.
Chen does not teach wherein the steps further include: determining whether the input time series is anomalous Internet of Things (loT) communications based on the comparing
However Pourhmadi teaches wherein the steps further include: determining whether the input time series is anomalous Internet of Things (loT) communications based on the comparing (Pourhmadi page 3 line 9-17, Nonetheless, while edge servers serve IoT devices, they can be subject to attacks originating from the latter. In fact, IoT devices extend the attack surface and put the network at risk. IoT devices are highly vulnerable to malware, allowing attackers to use them as part of a botnet in order to launch different kinds of attacks toward the edge of the network. The security of the network edge is not only threatened by vulnerable IoT devices but is also subject to attacks targeting every asset of the edge, including its infrastructure and the virtualization technologies running on top of edge servers, such as those related to hypervisors, virtual machines and containers.
Page 6 line 22-31, Hence, in order to reduce the false positive rate of unsupervised learning techniques without falling into a fully supervised approach hindering the generalization of the solution to unknown anomalies, semi-supervised technique leveraging a few anomalous labelled data has been used. The common anomaly-based approaches ( e.g., neural networks) learn the distribution of the normal data. Then, the trained model is used to assign anomaly scores to test inputs such that it assigns high score to Out-Of-Distribution (OOD) samples (i.e., the samples that deviate from the distribution of data which was used to train the model). To enhance the detection accuracy, the model may be trained on few OOD inputs as well to better detect and expose them. This technique is referred to as Outlier Exposure (OE)
Figure 1
PNG
media_image14.png
728
556
media_image14.png
Greyscale
[determining whether the input time series is anomalous Internet of Things (loT) communications].
Page 48 line 13-22,
An example of the online anomaly detection process is represented in FIG. 11. Download of FL anomaly detection model from FL server 17 (Block S 158). Collection of network traffic (Block S 160). Extraction of the defined features of network traffic (Block S 162). Input of network traffic samples extracted features to the FL anomaly detection model (Block S164).
Calculation of the reconstruction error d(Xtest) (Block S 166). A determination is performed whether d(Xtest) is greater than a threshold (Block S 168). If d(Xtest) is greater than a threshold, flagging of sample of anomalous is performed (Block S 170). If d(Xtest) is less than a threshold, flagging of sample as benign is performed (Block S172) [based on the comparing].).
Chen and Pourahmadi are considered to be analogous to the claim invention because they are in the same field of machine learning for anomaly detection. Therefore, it would have been obvious to someone of ordinary skill in the art before the effective filling date of the claimed invention to have modified Chen to incorporate the teachings of Pourahmadi to detect out-of-distribution data. Doing so to better classify computing systems network traffic and providing a defense against attacks and anomalies (Pourahmadi page 7 line 15-25, An attack on an edge server may not only disrupt the services provided by that edge but can also escalate and hinders the security of other edges and the central cloud as well. Further, an edge server can experience attacks that were not encountered hy another edge.
Hence, there is a need for a distributed anomaly detection solution that fulfills the following requirements
Can leverage the knowledge gained from anomalies at different edges.
Preserves users' privacy.
Can cope with the sheer and variable amount of data across the edges.
Leverages collaboration between the edges for an increased detection accuracy.
Page 6 line 26-32 and 7 line 1, In one or more embodiments described herein, an anomaly detection solution is provided that satisfies the above requirements. The anomaly detection solution provides edge severs with a defense mechanism against different types of attacks and anomalies. The one or more embodiments leverage FL to train a ML model, e.g., an Autoencoder, augmented with OE capability. The application of FL to the anomaly detection solution enables different edges, acting as FL clients, to collaborate under the orchestration of a central node (i.e., FL server) in order to better classify network traffic samples.).
Regarding claim 10 and analogous claim 20, Chen in view of Graham teach the method of claim 1 and analogous claim 11.
Chen and Graham are combine in the same rational as set forth above with respect to claim 1 and analogous claim 11.
Chen and Pourhmadi are combine in the same rational as set forth above with respect to claim 9 and analogous claim 19.
Pourhmadi teaches wherein the steps further include: determining whether the input time series corresponds to a Distributed Denial of Service (DDoS) network flow based on the comparing (Pourhmadi Page 6 line 22-31, Hence, in order to reduce the false positive rate of unsupervised learning techniques without falling into a fully supervised approach hindering the generalization of the solution to unknown anomalies, semi-supervised technique leveraging a few anomalous labelled data has been used. The common anomaly-based approaches ( e.g., neural networks) learn the distribution of the normal data. Then, the trained model is used to assign anomaly scores to test inputs such that it assigns high score to Out-Of-Distribution (OOD) samples (i.e., the samples that deviate from the distribution of data which was used to train the model). To enhance the detection accuracy, the model may be trained on few OOD inputs as well to better detect and expose them. This technique is referred to as Outlier Exposure (OE)
Page 35 line 20-23, According to one or more embodiments, the unlabeled network data is unlabeled benign network data, and the labeled network attack data is limited labeled attack data. According to one or more embodiments, the anomaly detection corresponds to performing distributed denial-of-service, DDoS, detection.
Page 48 line 13-22,
An example of the online anomaly detection process is represented in FIG. 11.
Download of FL anomaly detection model from FL server 17 (Block S 158). Collection of network traffic (Block S 160). Extraction of the defined features of network traffic (Block S 162). Input of network traffic samples extracted features to the FL anomaly detection model (Block S164).
Calculation of the reconstruction error d(Xtest) (Block S 166). A determination is performed whether d(Xtest) is greater than a threshold (Block S 168). If d(Xtest) is greater than a threshold, flagging of sample of anomalous is performed (Block S 170). If d(Xtest) is less than a threshold, flagging of sample as benign is performed (Block S172) [determining whether the input time series corresponds to a Distributed Denial of Service (DDoS) network flow based on the comparing].).
Pertinent Prior Art
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure.
Liu, Zhenzhen, et al. "Unsupervised out-of-distribution detection with diffusion inpainting." International Conference on Machine Learning. PMLR, 2023 – teaches a method for determining a distance from in-distribution samples and out-of-distribution samples.
Ramakrishna, Shreyas, et al. "Efficient out-of-distribution detection using latent space of β-vae for cyber-physical systems." ACM Transactions on Cyber-Physical Systems (TCPS) 6.2 (2022): 1-34 – teaches a method for out-of-distribution detection.
Conclusion
Any inquiry concerning this communication or earlier communications from the examiner should be directed to ALFREDO CAMPOS whose telephone number is (571)272-4504. The examiner can normally be reached 7:00 - 4:00 pm M - F.
Examiner interviews are available via telephone, in-person, and video conferencing using a USPTO supplied web-based collaboration tool. To schedule an interview, applicant is encouraged to use the USPTO Automated Interview Request (AIR) at http://www.uspto.gov/interviewpractice.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, Michael J. Huntley can be reached at (303) 297-4307. The fax phone number for the organization where this application or proceeding is assigned is 571-273-8300.
Information regarding the status of published or unpublished applications may be obtained from Patent Center. Unpublished application information in Patent Center is available to registered users. To file and manage patent submissions in Patent Center, visit: https://patentcenter.uspto.gov. Visit https://www.uspto.gov/patents/apply/patent-center for more information about Patent Center and https://www.uspto.gov/patents/docx for information about filing in DOCX format. For additional questions, contact the Electronic Business Center (EBC) at 866-217-9197 (toll-free). If you would like assistance from a USPTO Customer Service Representative, call 800-786-9199 (IN USA OR CANADA) or 571-272-1000.
/ALFREDO CAMPOS/Examiner, Art Unit 2129
/HAL SCHNEE/Primary Examiner, Art Unit 2129
Prosecution Insights