From Pixels to Privacy: Temporally Consistent Video Anonymization via Token Pruning for Privacy Preserving Action Recognition

Dimensions overall score 7.0

• The proposed attention-driven spatiotemporal video anonymization framework uses two task-specific classification tokens (action CLS and privacy CLS) within a shared Transformer backbone to disentangle utility and privacy features.method 95%
  Evidencepartial

  we introduce two task-specific classification tokens, an action CLS token and a privacy CLS token, that learn complementary representations within a shared Transformer backbone.

  Implicationpartial

  This is a core technical detail of the proposed method, explicitly described in the abstract and detailed in the method section.

  Verificationpartial

  partial

• The framework computes a utility-privacy score for each spatiotemporal tubelet by contrasting attention distributions of the action and privacy CLS tokens, and prunes tubelets with the highest scores.method 95%
  Evidencepartial

  We contrast their attention distributions to compute a utility-privacy score for each spatiotemporal tubelet, and keep the top-k tubelets with the highest scores.

  Implicationpartial

  This describes the key mechanism for anonymization, directly explained in the abstract and method section.

  Verificationpartial

  partial

• The proposed approach maintains action recognition performance comparable to models trained on raw videos while substantially reducing privacy leakage.result 90%
  Evidencepartial

  Extensive experiments demonstrate that our approach maintains action recognition performance comparable to models trained on raw videos, while substantially reducing privacy leakage.

  Implicationpartial

  This is a primary result claimed in the abstract and supported by comparative tables showing performance metrics.

  Verificationpartial

  partial

• On the VPUCF dataset, the proposed method achieves 80.96% Top-1 accuracy for action recognition with a 60.30% reduction in privacy leakage (cMAP).result 90%
  Evidencepartial

  Ours 80.96 (↓3.92)60.30(↓16.32)0.531(↓0.153)

  Implicationpartial

  Specific numerical results are provided in Table 1, directly supporting this claim.

  Verificationpartial

  partial

• On the VPHMDB dataset, the proposed method achieves 79.59% Top-1 accuracy for action recognition with a 70.32% reduction in privacy leakage (cMAP).result 90%
  Evidencepartial

  Ours 80.96 (↓3.92)60.30(↓16.32)0.531(↓0.153)79.59(↓1.46) 70.32 (↓6.3) 0.562 (↓0.122)

  Implicationpartial

  Specific numerical results are provided in Table 1, directly supporting this claim.

  Verificationpartial

  partial

• The VPUCF and VPHMDB datasets are suitable for evaluating privacy-preserving video analysis methods as they are annotated for both action recognition and five privacy-related attributes (face, skin color, gender, nudity, familiar relationships).technical 95%
  Evidencepartial

  VPUCFandVPHMDB[17] are large-scale video datasets annotated for both action recognition and privacy attributes. Both datasets provide video-level annotations for five privacy-related attributes:face,skin color,gender,nudity, andfamiliar relationships, making them ideal for assessing privacy-preserving video analysis methods.

  Implicationpartial

  The paper explicitly describes these datasets and their annotations, highlighting their suitability for the research.

  Verificationpartial

  partial

• The method is evaluated on its ability to generalize to novel actions and novel private attributes, showing competitive performance compared to STPrivacy.result 85%
  Evidencepartial

  Table 2. Comparison of different privacy-preserving methods onnovel actionsandnovel privateattributes datasets.

  Implicationpartial

  Table 2 presents results for novel actions and attributes, comparing the proposed method to a baseline, indicating this evaluation was performed.

  Verificationpartial

  partial

• The proposed framework uses two task-specific classification tokens, an action CLS token and a privacy CLS token, within a shared Transformer backbone to disentangle utility and privacy features.method 95%
  Evidencepartial

  we introduce two task-specific classification tokens, an action CLS token and a privacy CLS token, that learn complementary representations within a shared Transformer backbone.

  Implicationpartial

  The abstract explicitly describes the core technical innovation of using two distinct CLS tokens for disentanglement.

  Verificationpartial

  partial

• The framework computes a utility-privacy score for each spatiotemporal tubelet by contrasting the attention distributions of the action and privacy CLS tokens.method 95%
  Evidencepartial

  We contrast their attention distributions to compute a utility-privacy score for each spatiotemporal tubelet, and keep the top-k tubelets with the highest scores.

  Implicationpartial

  The abstract clearly explains the mechanism for scoring tubelets based on attention distributions.

  Verificationpartial

  partial

• The proposed method maintains action recognition performance comparable to models trained on raw videos while substantially reducing privacy leakage.result 90%
  Evidencepartial

  Extensive experiments demonstrate that our approach maintains action recognition performance comparable to models trained on raw videos, while substantially reducing privacy leakage.

  Implicationpartial

  The abstract states this as a key finding, and Table 1 provides supporting numerical evidence showing competitive action recognition (Top-1) and improved privacy metrics (cMAP, F1) compared to raw data and other methods.

  Verificationpartial

  partial

• The framework selectively prunes spatiotemporal tubelets dominated by privacy cues while preserving those critical for action recognition.method 95%
  Evidencepartial

  This selectively prunes tubelets dominated by privacy cues while preserving those most critical for action recognition.

  Implicationpartial

  This is a direct consequence of the scoring and pruning mechanism described in the abstract and method section.

  Verificationpartial

  partial

• The VPUCF and VPHMDB datasets are used for evaluation and are annotated for both action recognition and five privacy-related attributes (face, skin color, gender, nudity, familiar relationships).technical 95%
  Evidencepartial

  VPUCF and VPHMDB[17] are large-scale video datasets annotated for both action recognition and privacy attributes. Both datasets provide video-level annotations for five privacy-related attributes:face,skin color,gender,nudity, and familiar relationships, making them ideal for assessing privacy-preserving video analysis methods.

  Implicationpartial

  The paper explicitly describes the datasets and their annotations, making them verifiable for evaluating the method.

  Verificationpartial

  partial

Competitive landscape

A novel framework for video anonymization that prunes privacy-sensitive information while preserving action recognition accuracy, enabling safer video analytics.