Fig. 1 Example multilevel attributes in a defect ontology characterizing two defects: layer misalignment and crack More about this image found in Example multilevel attributes in a defect ontology characterizing two defec...

Fig. 2 Decisions involved in the proposed DP-attribute learning framework under the context of federated learning. Four decisions are outlined: selection of ontology branch depth to share (thick arrows on the left running through ontology) and federating strategy (thick arrows on the right selecti... More about this image found in Decisions involved in the proposed DP-attribute learning framework under th...

Fig. 3 DP embedding for attribute learning More about this image found in DP embedding for attribute learning

Fig. 4 Federated DP-SGD for attribute learning More about this image found in Federated DP-SGD for attribute learning

Figure 0015 More about this image found in Untitled image

Fig. 5 Examples of training data (seen classes) and testing data (unseen classes) in the dataset More about this image found in Examples of training data (seen classes) and testing data (unseen classes) ...

Fig. 6 Availability of image samples for training provided by Manufacturer 1 and Manufacturer 2. Defects for which manufacturers did not provide morphology data were used as testing data. More about this image found in Availability of image samples for training provided by Manufacturer 1 and M...

Fig. 7 Available ontology for Manufacturer 1 and Manufacturer 2 More about this image found in Available ontology for Manufacturer 1 and Manufacturer 2

Fig. 8 Selected optimal ontology (dark gray cells) for Manufacturer 1 and Manufacturer 2 More about this image found in Selected optimal ontology (dark gray cells) for Manufacturer 1 and Manufact...

Fig. 9 The confusion matrix for Manufacturer 1 and Manufacturer 2 based on the ontology structure of Fig. 8 More about this image found in The confusion matrix for Manufacturer 1 and Manufacturer 2 based on the ont...

Fig. 10 The confusion matrix for Manufacturer 1 based on the expanded test results with ten testing classes, include a mixture of seen and unseen classes. Test 5, Test 6, and Test 8 are unseen classes. More about this image found in The confusion matrix for Manufacturer 1 based on the expanded test results ...

Fig. 11 Zero-shot defect classification accuracy More about this image found in Zero-shot defect classification accuracy

Fig. 12 Privacy expenditure and the number of unseen classes in zero-shot learning More about this image found in Privacy expenditure and the number of unseen classes in zero-shot learning

Fig. 13 The median plot of privacy spent against the number of federating rounds N and upper-bound budget for Manufacturer 1 and Manufacturer 2. The privacy spent generally increases as the number of federating rounds increases ( N = 5 to N = 30 ). Within each federating round,... More about this image found in The median plot of privacy spent against the number of federating rounds ...

Fig. 14 The median of privacy spent against the number of federating rounds N and upper-bound budget for Manufacturer 1 and Manufacturer 2. For high upper-bound limits, the privacy spent increases as the number of federating rounds increases. On the other hand, the privacy spent is limited a... More about this image found in The median of privacy spent against the number of federating rounds N ...