The present article, on the one hand, addresses issues remained open in [9], in particular an experimental validation, and, on the other hand, provides the design of a new generic attack that combines the Pixels of Interest (PoI) strategy with the Noise Blowing Up (NBU) method. The resulting PoI+NBU method aims at enhancing the effectiveness of any type of attack (white-box or black- box) and of any specific attack on CNNs at the creation of HR adversarial images of exceptional visual quality.

This combination works as follows in practice. A clean high-resolution image is reduced to the LR domain to fit the input size of a CNN to attack. The PoI strategy is applied in the LR domain to identify the most relevant areas of an image for its classification by the considered CNN. Then an attack is performed, focusing on these zones, thereby reducing its search space and enhancing its efficiency. The adversarial noise, created that way in limited zones in the LR domain, is blown-up to the HR domain. This noise is then added to the HR clean image, leading to a high-resolution adversarial image, indistinguishable from the original HR clean image for a human eye.

We validate the combined PoI+NBU approach experimentally. Specifically, we employ a variant of the evolutionary algorithm-based (EA) attack described in [10] on 100 high resolution (HR) clean images, targeting the MobileNet CNN [11] trained on ImageNet.

Section 2 outlines the key theoretical steps of the PoI+NBU strategy. Section 3 lists the targeted CNN, the HR clean images, and the essential features of the EA-based targeted attack used in the experiments. Section 4 presents the outcome of the experiments, including a visual assessment of the quality of the adversarial images obtained through some illustrative images. Section 5 summarizes the findings of this paper.

The algorithms and experiments were implemented in Python 3.9 utilizing the NumPy 1.23.5, TensorFlow 2.14.0, Keras 3, and Scikit 0.22 libraries. Computational tasks were executed on nodes equipped with Nvidia Tesla V100 GPUs within the IRIS HPC Cluster at the University of Luxembourg [12]. Additional material (clean images used, their size, example of adversarial images, and source code) can be retrieved at https://github.com/emancellari/PoI_NBU.git

Attacks are classified according to the level of knowledge an attacker has about the CNN to deceive. In white-box attacks [13, 14, 15], the attacker has complete knowledge of the target CNN’s architecture, parameters, and training data, allowing for precise creation of adversarial images, often with high success rates. In contrast, black-box attacks [10, 16, 17, 18] rely only on observing the input-output behavior of the target model, typically requiring more time and resources.

Attack scenarios are manifold. Given a clean image classified by the CNN in a category c_a, in the target scenario, one selects a category c_t ≠ c_a , and one adds adversarial noise to the clean image to create an adversarial image classified by the CNN in c_t. As such, one has defined a good enough adversarial image. A τ -strong adversarial image (for 0 < τ ≤ 1) is an adversarial image classified in c_t with a c_t-label value ≥ τ . In the untargeted scenario, the process is similar as in the target scenario, except that one requires the adversarial image to be classified in any category c ≠ c_a .

Finally, adversarial images can be indistinguishable for a human as compared to the associated clean images, or not. The former requirement is clearly much more challenging than the latter one.

Pixels of Interest (PoI) strategy

Figure 1 describes the PoI process in the LR domain. One is given a CNN C to deceive, and a clean image A, of size equal to the input size of C (say 224 x 224 if C is trained on ImageNet), classified by C as belonging to the category c_a with c_a-label value equal to τ_a.

One uses BagNet [19] to identify the pixels relevant for a CNN’s classification of the image in c_a (b) and in c_t (d), thanks to a heatmap. Note that one does not specify which CNN we are dealing with, so that making use of BagNet is compliant with the requirements set by black-box attacks. Then we sieve these pixels and keep only the x% most significant for c_a on the one hand and for ct on the other hand, where x is fixed at will ((c) and (e)). One merges this information (without redundancy) in (f). The attack is performed on these pixels of interest, leading to an adversarial image classified by C in the target category c_t with a c_t-label value equal to _t.

FIGURE 1
PoI process in the LR domain for any attack, any scenario, and any CNN

FIGURE 1. PoI process in the LR domain for any attack, any scenario, and any CNN.

Remarks. Firstly, one could use clustering techniques like DBSCAN [20] to encapsulate these top x% most relevant pixels into larger zones of interest before performing the attack. Doing so presents the advantage of a lesser concentration of the attack on individual pixels, what may lead to a better visual quality. However, it does not prove true in practice (essentially because an observer notices rectangles on the adversarial images obtained). Moreover, it often implies that very large proportions of the image are subject of the attack, even if one uses only the top 1% most relevant pixels: our experiments showed that one jumps from 4.70% of the image without DBSCAN (see Table 2 and Figure 3 in additional material file), to 65% with DBSCAN, thereby adding a very large proportion of less-relevant pixels to the attack, leading to a slowing down of the process and lesser success rates. In other words, clustering techniques are unlikely to provide any substantial advantage.

Secondly, BagNet acts as a proxy of the CNN to attack but does not substitute it. Therefore, the usage of BagNet is compatible with a black-box attack scheme.

Thirdly, one can see our PoI strategy as a generalisation of the attacks [21, 22], where one or a few pixels are modified to create adversarial images. However, our aim goes beyond, since, as opposed to the aforementioned attacks where a human immediately sees that an attack occurred, we intend to create adversarial images indistinguishable from the original clean image.

In a nutshell, in the Noise Blowing-Up (NBU) generic strategy [8] illustrated in Figure 2, a clean HR image is reduced with a resizing interpolation function to fit the CNN C’s input size. c_a denotes the category in which C classifies this resized clean image. Then an attack atk is performed in the LR domain on this image to create an adversarial image classified in c ≠ c_a (which may be a predefined category ct in the target scenario). The adversarial noise is extracted in the LR domain and then blown-up to the HR domain to fit the original clean image size. This blown-up noise is then added to the HR clean image, leading to a HR tentative adversarial image. This image is again processed to fit C’s input size. If C classifies it in c, one has obtained that way a HR adversarial image.

FIGURE 2
The Noise Blowing-Up strategy

FIGURE 2. The Noise Blowing-Up strategy

The PoI+NBU method illustrated in Figure 3 integrates the PoI strategy with the NBU method to create high-resolution adversarial images effectively. The PoI strategy initially identifies the relevant regions of the resized clean image in the LR domain on which the attack will occur. Once the attack is applied within these zones, the NBU strategy is used to blow up the obtained adversarial noise to the HR domain, and the process continues as in Subsection 2.3.

A key advantage of this approach combining two generic strategies is that the result is again a generic strategy: It applies a priori to any attack, any scenario, and any CNN, and it is still a black-box attack.

For attacks that incorporate randomness, such as evolutionary-based attacks, rather than relying on a single substantial attack round which would create a very strong adversarial noise at once, one could also consider performing multiple rounds of moderate attacks, each leading to the creation of moderate noise [9], where for instance each round

FIGURE 3
The combined PoI+NBU strategy

generates focused adversarial noise within some particular zone of interest. Although none of them would be enough to create a HR adversarial noise, their collaborative efforts may. The successive layers of moderate noise, blown-up and carefully combined, may progressively generate an adversarial image in the HR domain.

The experimentation is performed for the target scenario for the 10 pairs (ca, ct) of cleantarget categories specified in Table 1 (same to those utilized in [10, 28])

TABLE 1. For 1 ≤ p ≤ 10, the 2nd row gives the ancestor category ca and its index number ap among the categories of ImageNet (Mutatis mutandis for the target categories, 3rd row).

TABLE 1
For 1 ≤p ≤ 10, the 2nd row gives the ancestor category ca and its index number ap among the categories of ImageNet (Mutatis mutandis for the target categories, 3rd row).

For each ancestor category ca, we picked at random 10 clean ancestor images from the ImageNet validation scheme in ca, provided that their sizes h x w satisfy h ≥ 224 and w ≥ 224. This ensures that these 100 clean images belong to the HR domain. Additional material (see end of Section 1) contains these images and their original sizes.

Once the pixels of interest are identified, one performs the black-box Evolutionary- based Algorithm (EA) attack [10] (see Algorithm 1 for its pseudo-code) within these regions, while keeping the rest of the pixels untouched.

The attack is executed for the target scenario to create 0.55-strong adversarial images (this ensures a convincing margin ≥ 0.10 with respect to the second best category). The maximum number of generations is set to N = 10, 000, and the population size is set to 40.

∈ controls the maximum allowable change in pixel values for the entire image, while α determines the magnitude of change for each pixel in each generation of the EA. These parameters play a crucial role in shaping the nature and magnitude of adversarial perturbations generated by the algorithm. Throughout the experiments, the value of α per generation is fixed at 1/255.

The EA is initially executed without PoI, with ∈ = 8, 12, and 16. Subsequently, it runs with PoI applied to increasing percentages of relevant pixels: the top x% (x: 10, 20, 25, 30, and 35) of both ca-label and ct-label values (taken together without any duplication) as measured by BagNet-33. This leads to 1800 attempts to generate adversarial images within the LR domain. Experience shows that the EA is unable to generate a significant number of adversarial images if x < 10, since in this case the proportion of the image affected is too narrow. Therefore, the study considers x ≥ 10.

Algorithm 1 EA attack pseudocode [10, 18]

1: Input: CNN C, initial image A, perturbation magnitude α, max perturbation ∈, ancestor class ca, target class index t, current generation g, max generations N

2: Initialize population: 40 copies of A; I0 as the first individual

3: Compute fitness for all individuals

4: while (OI0 [t] < τ ) & g < N do

5: Rank individuals by fitness: top 10 as elite, next 20 as middle class, last 10 as lower class

6: Mutate a random number of pixels in middle and lower class individuals withα; clip mutations to [ -∈ , ∈]

7: Replace lower class with mutated elite and middle class individuals

8: Cross-over individuals to form new population

9: Compute fitness for all individuals

The feasibility study is performed using MobileNet [11] trained on ImageNet [4]. We selected this CNN because it is optimized (and favored over other CNNs) for applications running on devices with limited processing power, memory, and storage capacity [29]. Examples of recent applications of MobileNet include the classification of freshwater fish on smartphones for farmers [30], the identification of tomato leaf disease in agriculture [31], the detection of skin cancer [32], etc.

Table 2 presents a comparison between MobileNet, the original GoogleNet [33] and VGG16 [34] in terms of the number of parameters, accuracy, and computational resources. MobileNet achieves nearly the same accuracy as VGG16 but with significantly smallersized parameters, being 32 times smaller, and requiring 27 times less computational resources (Mult-Adds). MobileNet outperforms GoogleNet in terms of accuracy while being smaller and requiring more than 2.5 times less computational resources.

TABLE 2. MobileNet vs original GoogleNet and VGG16: Details include parameter counts, ImageNet accuracy, and Mult-Adds (M-millions)

TABLE 2
MobileNet vs original GoogleNet and VGG16: Details include parameter counts, ImageNet accuracy, and Mult-Adds (M-millions)

TABLE 3. Average number of generations required to generate adversarial images (from acorn1 and maraca2) in the LR domain by EA without and with PoI guidance. Results are for ∈ = 8, 12, 16 and the top x% most relevant pixels for x = 10, 20, 25, 30, 35. The speed change is also given in percentages; negative values indicate a slower performance, and positive values a faster performance.

TABLE 3
Average number of generations required to generate adversarial images (from acorn1 and maraca2) in the LR domain by EA without and with PoI guidance. Results are for ∈ = 8, 12, 16 and the top x% most relevant pixels for x = 10, 20, 25, 30, 35. The speed change is also given in percentages; negative values indicate a slower performance, and positive values a faster performance.

Table 3 presents the performance of the EA in generating adversarial images in the LR domain, measured by the number of generations, both with and without PoI guidance. The results are based on acorn1 and maraca2 (see Additional material), as the EA successfully generated 0.55-strong adversarial images from these two clean images for all the mentioned settings (top x% and ), both with and without additional PoI guidance. The values are averaged for these two attempts.

When ∈ is increased, the performance of EA increases for all the top x% values. The best performance, in terms of the number of generations, of the EA with PoI is obtained when the top 35% of relevant pixels are used with ∈ = 16. It results in a 67.2% speed increase compared to the EA without PoI guidance. For ∈ = 16, Figure 4 shows how EA converges to the target category without PoI on the one hand, and with PoI using top 35% of the most relevant pixels for the (acorn1-rhinoceros beetle) ancestor-target pair on the other hand. EA’s learning period is drastically shortened when one uses PoI. Indeed, using PoI, the EA finds the path to the target category almost 60% faster than without PoI. This acceleration behavior is consistent across most of the ancestor-target pairs.

Figure 4
With ∈ = 16, EA’s convergence pattern from the clean category c_a towards the target category c_t without PoI (EA) and with PoI (EA_PoI, using the top 35% most relevant pixels) is analyzed for ca = _acorn1, c_t = rhinoceros beetle

Figure 5 illustrates, with the clean image acorn1, the visual quality of low-resolution adversarial images generated by the EA alone (without PoI), and when the EA is guided with PoI using the top 35% of relevant pixels for ∈ = 8, 12, 16. Results for other top x% are provided in Figure 2 of the Additional material. For a human, all obtained adversarial images are challenging to distinguish from the clean image.

In view of what precedes (in terms of speed and visual quality of adversarial images in the LR domain), we used = 16 and the top 35% most significant pixels identified by BagNet-33 for the remaining experiments combining PoI and NBU.

Using these parameters, the EA generated 56 0.55-strong adversarial images in the LR domain from 100 clean images. Out of the 56, NBU successfully converted 44 of them into HR adversarial images that MobileNet classifies in the target category for the (c_a, c_t) pair and target scenario specified in Table 1. Table 4 summarizes the results for these 44 HR adversarial images; numerical values are averaged. Its first column lists the clean image categories. Note that the brown_bear is not included because no 0.55-strong adversarial images were generated from this category. The second column shows the proportion of the image space that is identified by considering the top 35% most relevant pixels. It shows that the EA attack will focus on 70.4% of the clean LR image on average. The third column shows the average number of generations required by the EA to create a 0.55-strong adversarial image in the LR domain (on average, each generation takes between 0.90 and 0.99 seconds). The fourth column gives the average value of t (which is necessarily ≥ 0.55). The fifth column provides the average ct-label value τ_t for degraded adversarial images, and the sixth column gives the resulting average loss L_C() = _t − τ_t, where is the clean HR image classified in the original category c_a.

FIGURE 5
Visual quality of the low-resolution adversarial images generated with EA alone and with PoI guidance (using the top 35% of the most relevant pixels) in the LR domain for epsilon values 8, 12, and 16

FIGURE 5. Visual quality of the low-resolution adversarial images generated with EA alone and with PoI guidance (using the top 35% of the most relevant pixels) in the LR domain for epsilon values 8, 12, and 16.

On average, the NBU process caused a 0.251 label value loss. Despite this loss, the created HR adversarial images remain adversarial, achieving an average ct -label value of 0.301.

Table 5 provides the execution time (in seconds) of the main steps of the combined PoI+NBU strategy performed on two representative examples: the largest clean image canoe4 (2448 x 3264) and the smallest one llama4 (253 x 380). Using BagNet-33 to find the top 35% most relevant pixels of ca-label and ct-label values (combined without any duplication) takes 4.38 seconds. The following step is the attack performed in the LR domain. Its timing varies from one method to another. The EA attack required 23 minutes for one image and 84 minutes for the other. The NBU process blowing up the adversarial noise from the LR domain to the HR domain and adding it to the clean HR image is the last step. It takes less than a second.

Altogether, the PoI+NBU strategy per se takes only around 5 seconds and remains completely marginal as compared to the time required by the attack (the EA attack in the present feasibility study). This outcome demonstrates the efficiency of the PoI+NBU approach in generating high-resolution adversarial images with minimal time overhead, apart from the chosen attack method (less than 1% overhead in the case of the EA attack).

TABLE 4. Average metrics (for top 35% and ∈ =16 ) for generating 0.55-strong adversarial images in the LR domain, including pixels of interest size (avgPoI), number of generations (avgGens), target label value before (avg_ t) and after NBU (avg_ τt), and loss (avg_L).

TABLE 4
Average metrics (for top 35% and ∈ =16 ) for generating 0.55-strong adversarial images in the LR domain, including pixels of interest size (avgPoI), number of generations (avgGens), target label value before (avg_ t) and after NBU (avg_ τt), and loss (avg_L).

TABLE 5. Time performance of PoI+NBU using the largest and smallest clean images. One uses = 16, the top 35% most relevant pixels identified by BagNet- 33, and the EA-based attack. Values are in seconds.

TABLE 5
Time performance of PoI+NBU using the largest and smallest clean images. One uses = 16, the top 35% most relevant pixels identified by BagNet- 33, and the EA-based attack. Values are in seconds.

The visual quality of high-resolution adversarial images generated by the PoI+ NBU strategy for the EA-based attack is assessed on three examples in Figure 6. Its 1st row displays the HR clean images, and its 2nd row their corresponding HR adversarial images. Their names and sizes are at the top of each figure. Despite the added adversarial perturbations, the visual differences between the clean and adversarial HR images are imperceptible to the human eye. To further substantiate this observation, we computed the Fréchet Inception Distance (FID) [35] between clean and adversarial HR images and obtained an average FID score of 54.5.

FIGURE 6
Visual comparison between HR clean (1st row) and adversarial (2nd row) images.

enea.mancellari@uni.lu