Communications of the ACM

FEBRUARY 2020 | VOL. 63 | NO. 2 | COMMUNICATIONS OF THE ACM 87 attacker suffers intolerable image corruption (PSNR = 13dB, CW-SSIM = 0.56, CIEDE2000 = 34) by combining same number of randomly selected frames (“Illumination intensity randomization” section).

For the dynamic scene, we set fintra = 1 kHz and finter = 300 Hz (Section 3. 1). From Figure 9, we can see the authorized user has much higher quality (PSNR = 30dB, CW-SSIM = 0.98 in average) compared with attacker (PSNR = 10dB, CW-SSIM = 0.6 in average). This can be seen by resulting image frames in Figure 8, where attacker suffers from both intra-frame and inter-frame stripes. Thus LiShield’s authorization scheme is effective in unblocking specific users while maintaining protection against attackers.

6. 3. Effectiveness of barcode embedding

We first determine general optimal parameters for LiShield’s barcode detector in Section 4, based on the following metrics. (i) False alarm rate. We run the detector on 200 images (random real-world scenes) and measure the probability that a barcode is detected from clean image. (ii) Detection rate. We embed monochrome barcodes with different f1 from 400 Hz to 10 kHz with 200 Hz switching frequency. For each f1, we embed three frequencies with ∆f = 200 Hz interval and capture 300 images with these barcodes over a benchmark scene to obtain detection rate. Considering the trade-off between false alarm and detection, we choose Tb = 0.05 to bound the false alarm rate below 5%, while ensuring around 90% detection rate for monochrome barcode (Figure 10).

Using the above configuration, we found the detec-
tion rate for RGB barcode is close to 100% with or with-
out manual exposure attack, while being slightly below
Two bonus effects from our RGB LED are observed: (i)
The structural distortion from the stripes disrupts the cam-
era’s auto-focus function, often making the captured scene
extremely blur. This is because under LiShield, contrast
of bands no longer depends on focusing accuracy, which
breaks the assumption of auto-focus mechanism. (ii) The
color bands also mislead the automatic white balance function
across all five different scenes, since the camera can no lon-
ger identify a clean region in the image to calibrate itself and
thus hesitates.

Impact on dynamic scenes. To create a dynamic scene, we use the motor to rotate the smartphone, creating relative motion at three different speeds ( 45, 100, and 145 degrees/ second). Our experiment shows the average CW-SSIM among all three speeds further decreases by 0.1, which indicates that dynamic scene experiences worse quality under LiShield due to motion blur. Moreover, if the exposure time is larger than 1/100 s, then overexposure and motion blurs together further reduce the quality (PSNR < 6, CW-SSIM < 0.1). Thus, dynamic objects further decrease the adjustment range of exposure time and make manual exposure attack more ineffective.

6. 2. Effectiveness of user authorization

We developed an app (Section 5) that allows a user to capture critical frames on static scene protected by our RGB LED and then recover the scene following Section 3. The resulting image quality (PSNR = 25dB, CW-SSIM = 0.9, CIEDE2000 = 5) is comparable to the ideal setting when we disable LiShield’s LED modulation (Figure 8 shows example frames extracted from a recorded video). In contrast, the