PTAB

IPR2025-01398

CrowdStrike Inc v. Skysong Innovations LLC

Log In
Key Events
Petition
petition

1. Case Identification

2. Patent Overview

  • Title: Computer-Implemented System for Improving Classification of Criminal Activities
  • Brief Description: The ’900 patent discloses a computer-implemented system for improving the classification of data from deep web or dark net environments. The system uses machine learning to automatically assign hierarchical labels (tags) to discussion topics and includes methods to ensure hierarchical consistency, such as adding parent tags or removing child tags based on prediction probabilities.

3. Grounds for Unpatentability

Ground 1: Obviousness over Deliu and Pendar - Claims 1, 2, 4, 5, 10, 12, and 13 are obvious over Deliu in view of Pendar.

  • Prior Art Relied Upon: Deliu (a 2017 doctoral thesis) and Pendar (Patent 9,330,167).
  • Core Argument for this Ground:
    • Prior Art Mapping: Petitioner argued that Deliu disclosed a computer-implemented system for classifying data from a deep web hacker forum (Nulled.IO) using machine learning classifiers (e.g., SVM, k-NN) and extracting features by assigning word vectors. However, Deliu did not explicitly teach a hierarchical tag structure or generating a prediction list with probability values for multi-label classification. Petitioner asserted that Pendar supplied these missing elements, teaching a system for classifying textual data using a predefined hierarchical taxonomy (e.g., food service categories) and assigning labels by comparing a calculated prediction score for each tag against a predetermined threshold. Pendar further taught adding all dominant parent tags for any assigned child tag to enforce the hierarchy.
    • Motivation to Combine: A Person of Ordinary Skill in the Art (POSITA) would combine Pendar's hierarchical classification techniques with Deliu's system to improve classification accuracy. Deliu’s own work acknowledged the need for more complex classification, and hierarchical classifiers like those in Pendar were known to provide more accurate and representative results for real-world data, which is often inherently hierarchical.
    • Expectation of Success: A POSITA would have a reasonable expectation of success in this combination. Modifying Deliu's classifiers to incorporate Pendar's well-known hierarchical labeling and multi-label scoring techniques involved applying predictable machine learning strategies to improve a known system.

Ground 2: Obviousness over Deliu, Pendar, and Frank/Ramírez-Corona - Claim 3 is obvious over Deliu, Pendar, and Frank or, alternatively, over Deliu, Pendar, and Ramírez-Corona.

  • Prior Art Relied Upon: Deliu, Pendar, Frank (a 2000 university thesis), and Ramírez-Corona (a 2014 journal article).
  • Core Argument for this Ground:
    • Prior Art Mapping: This ground depended on the construction of claim 3, which contained a typographical error (see Section 4).
      • With Correction (Frank): Assuming the claim was corrected to require removing child tags when a parent tag's probability is below a threshold, Petitioner argued Frank taught this limitation. Frank disclosed a "critical value pruning" technique for decision trees where a subtree (i.e., child tags) is removed if the value at the corresponding node (i.e., parent tag) is below a fixed threshold.
      • Without Correction (Ramírez-Corona): If the claim was read as written (requiring removal based on a child tag's probability), Petitioner argued Ramírez-Corona taught this. Ramírez-Corona disclosed a bottom-up pruning method where traversal up a hierarchy is stopped and descendants (child tags) are pruned if the "probability of the most probable child" is less than a threshold.
    • Motivation to Combine: A POSITA would be motivated to add a pruning step to the Deliu/Pendar system to improve classifier performance. Removing lower-level tags that contradict the natural hierarchy (e.g., have low probability scores) reduces false positives and enforces consistency, a known benefit of pruning techniques like those taught by Frank and Ramírez-Corona.
    • Expectation of Success: A POSITA would expect success in applying these known pruning methods to the combined Deliu/Pendar system, as it would leverage the existing probability scores to implement a standard technique for improving hierarchical classification.

Ground 3: Obviousness over Deliu, Pendar, and Chawla - Claims 6 and 14 are obvious over Deliu, Pendar, and Chawla.

  • Prior Art Relied Upon: Deliu, Pendar, and Chawla (a 2002 journal article).

  • Core Argument for this Ground:

    • Prior Art Mapping: Petitioner asserted that the base system was taught by Deliu and Pendar. Chawla was introduced to teach the limitations of claims 6 and 14, which relate to addressing class imbalance in a dataset. Chawla disclosed the Synthetic Minority Over-sampling Technique (SMOTE), a method for creating synthetic data samples to add to a ground truth dataset. This is done by creating sample points along a line defined between minority feature vectors, which Petitioner argued was a near-verbatim disclosure of the process described in the ’900 patent and recited in claim 6.
    • Motivation to Combine: Deliu noted that its dataset could have a limited number of samples, a problem that can lead to class imbalance. A POSITA would be motivated to implement Chawla's well-known SMOTE technique to overcome this known limitation in machine learning. This would improve the accuracy of Deliu's classifiers by providing more minority class samples for the model to learn from.
    • Expectation of Success: Implementing SMOTE would be a straightforward and predictable modification. Chawla provided pseudo-code for the technique, and its benefits for improving classifier performance on imbalanced datasets were well-documented.

  • Additional Grounds: Petitioner asserted additional obviousness challenges based on Deliu and Pendar in combination with Xiao (for using an elastic search similarity score to supplement a dataset), Ukrainczyk (for data tokenization and stemming/lemmatization), and Nunes (for using a web crawler to retrieve deep web forum data).

4. Key Claim Construction Positions

  • Petitioner argued that Claims 3 and 4 contain obvious typographical errors that should be corrected for the proceeding.
  • Claim 3: As written, the claim required removing child tags based on a probability value associated with "the child tag," which lacked an antecedent basis. Petitioner argued, based on the specification's "Algorithm 1" and the prosecution history, that the claim should be construed to mean removing child tags where the probability value associated with the corresponding parent tag is below a threshold.
  • Claim 4: Similarly, Petitioner argued that claim 4 should be corrected to require adding parent tags based on the probability of the child tag, not the "parent tag" as written, to align with the specification and prosecution history.

5. Relief Requested

  • Petitioner requests institution of an inter partes review and cancellation of claims 1-14 of the ’900 patent as unpatentable.