Multimedia Forensics

Question 1

Reorder in the correct sequence the following operations of the JPEG processing chain.

DCT
Entropy coding
Quantization
Block partitioning
Color conversion RGB to YUV

Answer 1

1) [Color conversion RGB to YUV]
2) [Block partitioning]
3) [DCT]
4) [Quantization]
5) [Entropy coding]

Question 2

Classify the following strategies into passive and active ones.

Watermarking 
CFA estimation
Steganography
PRNU estimation 
Leveraging compression traces 
Encryption
Checking inconsistencies

Answer 2

Watermarking [Active]
CFA estimation [Passive]
Steganography [Active]
PRNU estimation  [Passive]
Leveraging compression traces [Passive]
Encryption [Active]
Checking inconsistencies [Passive]

Question 3

Given the following statements, assign the strategy (watermarking/cryptography) that fits the definition most.

1. It does not prevent from making untraceable copies ____
2. It allows a free distribution of the image/video content ______
3. It is unperceptible ____
4. it prevents unauthorized users from accessing the multimedia content _____
5. it can be applied multiple times without compromising the accessibility of the content

Answer 3

1. It does not prevent from making untraceable copies [cryptography]
2. It allows a free distribution of the image/video content [watermarking]
3. It is unperceptible [watermarking]
4. it prevents unauthorized users from accessing the multimedia content [cryptography]
5. it can be applied multiple times without compromising the accessibility of the content [watermarking]b

Question 4

Performances of a watermarking strategy can be measured using _______, while the performance of a data hiding strategy must be measured using _______. In both strategies it is possible to take into account ______ (or other quality measurements) and _______ to evaluate the impact of each approach on the rate-distortion performance.

Assign these:
ROC
PSNR 
BER
bit rate

Answer 4

Performances of a watermarking strategy can be measured using [ROC], while the performance of a data hiding strategy must be measured using [BER]. In both strategies it is possible to take into account [PSNR] (or other quality measurements) and [bit rate] to evaluate the impact of each approach on the rate-distortion performance.

Question 5

Complete the following sentences concerning the relation between forensic and antiforensic strategies.

1. Non-reversible processing operations
2. Antiforensic strategies want to
3. Forensic analysts

a. leave (unintentionally) some footprints on multimedia contents
b. cancel processing footprints left on the multimedia content
c. look for processing traces lefet by the editing operations

Answer 5

1. Non-reversible processing operations → leave (unintentionally) some footprints on multimedia contents,
2. Antiforensic strategies want to → cancel processing footprints left on the multimedia content.,
3. Forensic analysts → look for processing traces lefet by the editing operations

Question 6

Assign the proper definition/description to each concept.

1. In watermarking
2. In attacks
3. In steganography

a. the hidden information is supposed to be erased
b. the hidden information is supposed to be detectable only for certain users
c. the hidden information is supposed to be detectable

Answer 6

1. In watermarking → the hidden information is supposed to be detectable,
2. In attacks → the hidden information is supposed to be erased,
3. In steganography → the hidden information is supposed to be detectable only for certain users

Question 7

Select the correct sentences about chromatic aberration among the following ones:

a. It is due to the fact that the lens makes different wavelengths converge to different points.
b. It depends on light refraction passing from air to lens material and viceversa
c. It is more evident on the center of the lens
d. It can be compensated by an additional lens

Answer 7

b. It depends on light refraction passing from air to lens material and vice-versa,
d. It can be compensated by an additional lens,
a. It is due to the fact that the lens makes different wavelengths converge to different points.

Question 8

Concerning CMOS imaging sensors, we can state that (select the correct ones):

Select one or more:

a. dedicated circuitry is used to control the reading
b. they are based on a shift register circuitry
c. they can be activated because of a blooming effect
d. they are made of 3 transistors
e. values are sequentially read according to a row-column ordering

Answer 8

a. dedicated circuitry is used to control the reading,
d. they are made of 3 transistors,
e. values are sequentially read according to a row-column ordering

Question 9

An imaging sensor matrix is made of CCD units. Concerning each unit, we can state (select the correct ones):
Select one or more:
a. the first capacitor is charged according to light intensity
b. reading is sequential
c. the final image can be distorted because of a rolling shutter effect
d. charges are transferred using a shift register circuitry

Answer 9

d. charges are transferred using a shift register circuitry,

a. the first capacitor is charged according to light intensity

Question 10

Concerning the Photo Response Non Uniformity (PRNU) noise, we can state that (select the correct ones):

Select one or more:

a. it can be easily estimated from a highly texture image
b. it is robust to compression and most editing steps
c. it generates a multiplicative noise components depending on light intensity
d. it depends on temperature
e. it is generated by artifacts in the fabrication process of the sensor matrix.

Answer 10

e. it is generated by artifacts in the fabrication process of the sensor matrix.,
c. it generates a multiplicative noise components depending on light intensity

Question 11

Given an image 𝐼 under examination and the PRNU pattern 𝐾̂ (associated to a camera), select the most appropriate operation to verify that the image 𝐼 has been taken with the camera associated to 𝐾̂

a. Compute the correlation between 𝐾̂ and 𝑊̂, where 𝑊̂ is the residual obtained by denoising 𝐼
b. Compute the correlation between 𝐼 and 𝐾̂
c. Compute the correlation between 𝐾̂ 𝐼 and 𝑊̂, where 𝑊̂ is the residual obtained by denoising 𝐼
d. Compute the correlation between 𝐼 and 𝑊̂, where 𝑊̂ is the residual obtained by denoising 𝐼

Answer 11

c. Compute the correlation between 𝐾̂ 𝐼 and 𝑊̂, where 𝑊̂ is the residual obtained by denoising 𝐼

Question 12

In order to verify the correpondence between a PRNU estimate and an image, which of the following metrics can be used?

Select one or more:

a. Peak correlation energy
b. Normalized Cross Correlation
c. Average ratio between PRNU noise and residual after denoising
d. MSE between residual and PRNU noise

Answer 12

b. Normalized Cross Correlation,

a. Peak correlation energy

Question 13

Given a digital image, estimating the CFA interpolation strategy adopted on that image allows …

Select one or more:

a. … to associate the image to a specific camera.
b. … to associate the image to a specific acquisition time.
c. … to associate the image to a camera model.
d. … to associate the image to a bayer pattern

Answer 13

c. … to associate the image to a camera model.

d. … to associate the image to a bayer pattern

Question 14

Among the following list, select the applications that can be enabled by multimedia forensics tools.

Select one or more:

a. Copyright violation detection
b. Fraud detection
c. 3D reconstruction
d. Image cyphering
e. image/video authentication

Answer 14

e. image/video authentication
b. Fraud detection
a. Copyright violation detection

Question 15

Which of the following features can be used to identify the ink-jet printer that generated a given document.

Select one or more:

a. Dot shapes
b. Periodic variation in the missing jets or paper-advance errors
c. Optical photoconductor (OPC) angular velocity
d. Polygon mirror wobble

Answer 15

a. Dot shapes

b. Periodic variation in the miss

Question 16

Concerning fragile watermarking, we can state that …

Select one:

a. … it is necessary to reveal the editing operations applied on an image.
b. … it is an active strategy
c. … it is designed to remain unaltered after an editing operation.
d. … it is necessarily designed to be detected by everyone.

Answer 16

b. … it is an active strategy

Question 17

Considering a video re-capture, select the correct statement among the following:

Select one:

a. Recapture produces blocking artifacts on the acquired image.
b. Artifacts are never periodic
c. Ghosting artifacts are present.
d. Ghosting artifacts can be modelled as a non-linear operation

Answer 17

c. Ghosting artifacts are present.

Question 18

Concerning a recaptured video using a handheld device, select the wrong statement about motion cues

Select one:

a. motion cues are highly uncorrelated.
b. motion cues are computed from local feature points
c. motion cues are highly correlated

Answer 18

c. motion cues are highly correlated

Question 19

Assuming that you want to improve the clarity and visibility of a face in a frame from a video surveillance sequence taken at night by a low resolution camera, what operation would you consider … (among the following)

Select one or more:

a. magnification
b. frame integration
c. gamma correction
d. blurring compensation
e. motion ehnancement
f. noise reduction

Answer 19

a. magnification
b. frame integration
c. gamma correction
f. noise reduction

Question 20

Reorder sequentially the different steps of a face recognition algorithm (listed at the bottom of the paragraph).

a. [Face processing]
b. [Face alignment and normalization]
c. [Feature extraction]
d. [Face classification/verification]
e. [Face localization]

Answer 20

A) [Face localization]
B) [Face alignment and normalization]
C) [Face processing]
D) [Feature extraction]
E) [Face classification/verification]

Question 21

Given the different types of attacks to a facial recognition (FR) system (reported at the end of the section), link each characteristics to the attack type that proves to be the most suitable.

• the attack requires a picture (digital or printed) of the person to be impersonated ______
• the attack is very difficult to be detected (even for a human supervisor) ____
• shadows and depth cues can be used to reveal this attack ___
• It is less frequent than the other two ____

Assign these:

Photo attack
Mask attack
Make‑up attack

Answer 21

• the attack requires a picture (digital or printed) of the person to be impersonated [Photo attack]
• the attack is very difficult to be detected (even for a human supervisor) [Make‑up attack]
• shadows and depth cues can be used to reveal this attack [Photo attack]
• It is less frequent than the other two [Mask attack]

Question 22

In face recognition, morphing operations are intended to …
Select one or more:
a. … compensate different head orientations and poses.
b. … compensate differences in facial expressions
c. … modify the positions of the 68 facial landmarks
d. … compensate illumination difference

Answer 22

b. … compensate differences in facial expressions

c. … modify the positions of the 68 facial landmarks

Question 23

In face recognition systems, facial landmarks are detected and moved to compensate differences in people expressions. Concerning the 68 dlib facial landmarks, we can state that …

Select one or more:

a. … they can be localized around the mouth
b. … they are localized in the middle of cheeks.
c. … they can be localized on the chin
d. … they are localized in the middle of the forehead.
e. … they can be localized on the eyebrows

Answer 23

a. … they can be localized around the mouth
c. … they can be localized on the chin
e. … they can be localized on the eyebrows

Question 24

Among the following factors, select those that can be included among the PIE factors

Select one or more:

a. few samples per person.
b. compression
c. image resolution
d. head pose
e. facial expression
f. closed-set limitations
g. light intensity

Answer 24

b. compression
c. image resolution
d. head pose
e. facial expression
g. light intensity

Question 25

Which among the following processing steps are usually adopted in a video codec?

Select one or more:

a. motion prediction
b. psycho-acoustic models
c. quantization
d. subband decomposition
e. transform-coding

Answer 25

a. motion prediction
c. quantization
e. transform-coding

Question 25

Which among the following processing steps are usually adopted in a video codec?

Select one or more:

a. motion prediction
b. psycho-acoustic models
c. quantization
d. subband decomposition
e. transform-coding

Answer 25

a. motion prediction
c. quantization
e. transform-coding

Question 26

An image sensor matrix is made of CMOS units. Concerning each unit, we can state (select the correct one):

a. Pixel intensities are read using a shift register circuitry
b. the first capacitor is charged according to light intensity
c. reading is instantaneous for all the pixels/sensors of the matrix
d. reading is sequential

Answer 26

d. reading is sequential

Question 27

Concerning the PRNU noise, select the wrong statement among the following

a. It can be easily estimated from highly-textured images
b. it can be estimated for each camera instance (and looks different)
c. it is compromised by compression
d. it is generated by artifacts in the fabrication process of the sensor matrix

Answer 27

a. It can be easily estimated from highly-textured images

Question 28

Given the following statements about a watermarking strategy, select the correct one

a. It doesn’t alter the image/video content(pixel values or a bit rate)
b. it can be applied multiple times without compromising the accessibility
c. it allows managing access rights
d. It is similar to cryptography

Answer 28

b. it can be applied multiple times without compromising the accessibility

Question 29

Among the following list, select the strategies that can be employed by multimedia forensics to enhance the quality of visual data

a. Sharpening, distortion correction, magnification
b. PRNU estimation, CFA interpolation strategy estimation, chromatic aberration estimation
c. Watermarking, stabilization, image phylogeny
d. deepfake detection, integrity verification, ghosting artifacts detection

Answer 29

a. Sharpening, distortion correction, magnification

Question 30

Select which of the following coding strategies are NOT adopted by the JPEG image coding standard

a. entropy coding
b. quantization
c. transform coding
d. motion estimation

Answer 30

d. motion estimation

Question 31

Considering lab experience number 2 (creation of a GAN), select the most appropriate order of operations for a single iteration/step of training phase

Operations are:

a. generated_images = generator(noise, training=True)
b. gen_loss = generator_loss(fake_output)
c. disc_loss = discriminator_loss(real_output, fake_output)
d. fake_output=discriminator(generated_images, training=True)
e. real_output=discriminator(images, training=True)

Select one:

1. A, then C, then D, then B, then E
2. A, then E, then D, then B, then C
3. D, then A, then C, then B, then E
4. A, then B, then E, then C, then D

Answer 31

2. A, then E, then D, then B, then C

Question 32

Which among the following printer-related features can not be used for the identification of an ink-jet printer

a. Polygon mirror wobble
b. periodic variation by missing jets
c. paper advancement errors
d. dot shapes

Answer 32

a. Polygon mirror wobble

Question 33

In PRNU estimaton, the correlation between K J and W can be measured with different functions. Select the wrong statement among the following ones

Select one:

a. PCE is more reliable than NCC
b. PCE is more computationally expensive than NCC
c. PCE is based on NCC
d. PCE is. to be used if you have less than 5 flat-field images for the estimation of K.

Answer 33

d. PCE is. to be used if you have less than 5 flat-field images for the estimation of K.

Question 34

In a digital camera, radial distortion can be modelled by a parametric equation that is equal to ___

Select one:

a. r_u = k_1r_d + k_2r^2_d
b. r_u = k_1r_d
c. r_u = r_d + k_1r^2_d+k_2r^4_d
d. r_u = r_d + k_1r^3_d+k_2*r^5_d

Answer 34

d. r_u = r_d + k_1r^3_d+k_2r^5_d

Question 35

Concerning the PRNU equation
I = g^y[I_0(1+K)+ O]^y +Q

which among the following statements is surely wrong

a. I_0 is the noise-free image
b. K is a multiplicative term
c. g is gain factor
d. O includes compression noise

Answer 35

d. O includes compression noise

Question 36

Which of the following color is used in JPEG coding

Select one

a. HSV
b. RGB
c. YCbCr
d. Grayscale

Answer 36

c. YCbCr

Question 37

Select the wrong statement concerning subband audion representation

a. It is used in MP3 coding
b. the final representation quality depends only on subband regardless of the input signal to be reproduced
c. the more subbands, the higher is the representation accuracy
d. a signal/image is represented by a sum of multiple simpler components

Answer 37

b. the final representation quality depends only on subband regardless of the input signal to be reproduced

Question 38

Concerning generalized Benford’s law for forgery detection,
Klog_10(1 + 1/a+m^b)

it’s possible to state stat (select correct answer)

a. the value m assumes values in the range [0, +infin]
b. b is the most significant digit of a set of numbers
c. it has been used in election fraud detection
d. it is verified for all probability distribution functions

Answer 38

c. it has been used in election fraud detection

Question 39

Describe different camera-related authentication footprints for images and videos. Define at which steps of the image acquisition pipeline they are introduced and the strategies that can be adopted to highlight/extract them.

Question 40

Camera ballistic: describe how PRNU can be extracted and used to identify a specific camera.

Question 41

Overviews the different acquisition based footprints for a digital cameras and how they can be used to authenticate (associated the image to a given device or model) an image.

Question 42

overview some possible image tampering detection strate- gies.