May_EDFA_Digital

edfas.org ELECTRONIC DEVICE FAILURE ANALYSIS | VOLUME 21 NO. 2 34 thismethod requires a large amount of hand-labeled data from a similar type of IC, extensive training, and manual tuning and optimization. With RE, virtually all information about the structures of the IC die is either completely unknown or obscure. Therefore, techniques based on prior knowledge of these structures have a major drawback. As a result, there has beenaneed toapply unsupervisedmachine learning algo- rithms, which are applicable even to low-quality images. Unfortunately, this unsupervised operation has not yet been addressed by the RE community. For an example of the difficulties involved in IC reverse engineering from an imaging point of view, the authors performed a case study on a simple IC, namely a smart card. A smart card is a device that contains an embedded microprocessor for transactingdata. The reasons for selec- ting a smart card include its inherent simplicity compared to today’s typical microprocessors and its common use in security systems for authentication. The images shown in this case study (Figs. 3-5) were acquiredwith the following parameters: Magnification was set to 100 µm, dwelling time was 32 µs/pixel, and pixel density and excitation voltage were 1024 × 1024 and 5 kV, respectively. Imaging took 2 hours and 45 minutes for the entire IC (chip size 1.5 × 1.5 mm) with a 130-nm node technology. Even though the entire IC was deprocessed, only three major layers of importance to RE are discussed here: doped silicon, polysilicon, and the first metal layer. An example of the doped layer from the IC is shown in Fig. 3. Two sub-regions are highlighted in the image to emphasize one of the noise sources, i.e., the topographical spread, as discussedearlier. The spreadhas a cumulative effect on certain regionsmaking them brighter than their counterparts. This noise source is present in other regions as well, butwitha lower intensity. Nevertheless, it cannot be detected unless the image is observed at a higher magnification. The polysilicon region is shown in Fig. 4. The spread in Fig. 4a is associated with an improper delayering. The remnants from the doped region in conjunction with the structures in the current layer result inappar- ent spread in the lateral direction, like the one observed in the doped layer in Fig. 3. In Fig. 4b, the spread associated with the material along with the inherent noise of the imaging modality of the SEM causes the polysilicon structures to appear merged. These merged structures in the polysilicon layer would cause the functionality of the RE’d product to change. Therefore, it is cru- cial to ensure the image is acquired with parameters thatmake it possible to visualize distinct structures. The metal layer shown in Fig. 5 is the easiest to image because metal structures have lower spread than non-metal ones. Hence, their edges are easier to distinguish andwell-segmented compared tonon-metal structures. However, as illustrated in the example shown in Fig. 5a and 5b, the images are still susceptible to delayering defects. Fig. 3 The doped region. The two highlighted regions (a and b) emphasize topographical spread noise. Fig. 4 The polysilicon region. The two highlighted areas show (a) improper delayering and (b) merged structures. (b) (a) (b) (a)