Adaptive Pixel Value Differencing (APVD)

• Adaptive Pixel Value Differencing (APVD) is a steganographic method that adapts payload capacity per pixel pair based on local image smoothness or texture.
• It computes the absolute difference of non-overlapping pixel pairs and maps these differences to predefined ranges to determine embeddable bit capacities.
• APVD developments include overflow/underflow-aware adjustments, keyed pseudorandom selection, and deep-learning techniques for joint steganalysis and payload recovery.

to=shell 天天中彩票提款 code: {"command":["bash","-lc","python - <<'PY'\nimport urllib.request, urllib.parse, xml.etree.ElementTree as ET\nids=['(Deb et al., 20 Nov 2025)','(Hosain et al., 8 Jul 2025)','(Mandal et al., 2012)']\nfor aid in ids:\n url=f'http://export.arxiv.org/api/query?id_list={aid}'\n try:\n data=urllib.request.urlopen(url, timeout=20).read().decode('utf-8')\n print(aid, data[:300].replace('\n',' '))\n except Exception as e:\n print('ERR', aid, e)\nPY"]} to=shell 大发官网 code: {"command":["bash","-lc","python - <<'PY'\nimport urllib.request\nurl='https://export.arxiv.org/api/query?id_list=([2511.16604](/papers/2511.16604))'\ntry:\n with urllib.request.urlopen(url, timeout=20) as r:\n d=r.read(400).decode('utf-8','ignore')\n print(d)\nexcept Exception as e:\n print(type(e), e)\nPY"]} to=shell 天天中彩票上 code: {"command":["bash","-lc","echo hello"]} Adaptive Pixel Value Differencing (APVD) is a spatial-domain image steganographic method that extends Pixel Value Differencing (PVD) by adapting payload allocation to local image variation. In its standard form, APVD processes non-overlapping pixel pairs in a grayscale image, computes the absolute inter-pixel difference, assigns that difference to a predefined range, and embeds a number of payload bits determined by the width of the selected range. Smaller differences correspond to smoother regions and therefore lower embedding capacity, whereas larger differences correspond to textured or edge regions and therefore higher capacity. Across the literature considered here, APVD appears in three closely related roles: as an overflow/underflow-aware refinement of PVD, as the basis of a keyed pseudorandom pixel-selection scheme designed to mitigate “unused blocks,” and as the target of a unified deep-learning framework for joint stego detection and payload recovery (Mandal et al., 2012, Hosain et al., 8 Jul 2025, Deb et al., 20 Nov 2025).

1. Core formulation and adaptive embedding rule

APVD is described as an evolution of the classic Pixel Value Differencing method. Given a grayscale cover image, it is traversed in raster order and partitioned into non-overlapping adjacent pixel pairs, written for example as $\langle p_1, p_2\rangle$ or $(P_1,P_2)$. For each pair, the fundamental statistic is the absolute difference

$d = |p_2-p_1|.$

The interval $[0,255]$ is partitioned into disjoint ranges $R_i=[L_i,U_i]$, with width $|R_i|=U_i-L_i+1$. The number of embeddable bits for a pair whose difference falls in $R_i$ is

$b = \left\lfloor \log_2 (U_i-L_i+1)\right\rfloor.$

The adaptive principle is explicit: smooth regions, characterized by small $d$, receive fewer embedded bits, while textured regions, characterized by large $d$, receive more. One formulation gives example ranges such as $(P_1,P_2)$0, $(P_1,P_2)$1, $(P_1,P_2)$2, so that smooth areas embed $(P_1,P_2)$3 bits per pair while busy areas embed up to $(P_1,P_2)$4 or $(P_1,P_2)$5 bits per pair (Deb et al., 20 Nov 2025). A commonly used table reported in a later APVD extension, attributed there to Luo et al., is $(P_1,P_2)$6, $(P_1,P_2)$7, $(P_1,P_2)$8, $(P_1,P_2)$9, $d = |p_2-p_1|.$0, and $d = |p_2-p_1|.$1, with corresponding capacities $d = |p_2-p_1|.$2, $d = |p_2-p_1|.$3, $d = |p_2-p_1|.$4, $d = |p_2-p_1|.$5, $d = |p_2-p_1|.$6, and $d = |p_2-p_1|.$7 bits (Hosain et al., 8 Jul 2025).

Payload embedding is realized by interpreting the next $d = |p_2-p_1|.$8 secret bits as an integer $d = |p_2-p_1|.$9, then choosing a new difference within the same range:

$[0,255]$0

This range-preserving remapping is the essential APVD operation. It couples capacity control to local image complexity and is intended to balance visual imperceptibility with payload capacity. The literature emphasizes that this balance is achieved by allocating smaller ranges in smooth regions and larger ranges in textured regions, rather than by using a fixed bit budget per pair.

2. Pair reconstruction, extraction, and overflow/underflow-aware APVD

Once $[0,255]$1 has been chosen, the stego pair must be reconstructed. One presentation keeps the pair sum $[0,255]$2 constant and solves

$[0,255]$3

If $[0,255]$4, one reported solution is

$[0,255]$5

with the roles swapped otherwise (Deb et al., 20 Nov 2025). A related formulation, used in the 2025 pseudorandom-selection paper, defines $[0,255]$6 and applies a half-split adjustment to the pair so as to realize the new difference while preserving the average as closely as possible (Hosain et al., 8 Jul 2025).

The extraction rule is the inverse of the embedding rule at the difference level. For each stego pair, the decoder computes the observed difference $[0,255]$7, locates the corresponding range $[0,255]$8, computes the number of bits $[0,255]$9, and recovers the embedded integer as $R_i=[L_i,U_i]$0. The binary representation of $R_i=[L_i,U_i]$1 then yields the recovered payload bits. In the formulation summarized for APVD reverse steganalysis, extraction is performed pairwise over the entire stego image, reconstructing a bit-stream $R_i=[L_i,U_i]$2 from the sequence of recovered integers (Deb et al., 20 Nov 2025).

A distinct APVD formulation was introduced by Mandal and Das to exclude overflow and underflow explicitly. In that scheme, standard PVD embedding is attempted first; if the tentative stego pixels remain within $R_i=[L_i,U_i]$3, the pair is accepted. If a boundary violation occurs, the encoder inspects the most significant bit of the payload chunk used in that pair. When that MSB is $R_i=[L_i,U_i]$4, it is discarded, the bit-length is reduced by one, and the pair is recomputed. If overflow still occurs, the algorithm switches to a “single-pixel adjust,” keeping one pixel unchanged and moving the other by $R_i=[L_i,U_i]$5 so as to realize the desired change in difference. An overhead flag $R_i=[L_i,U_i]$6 records whether a bit was discarded, and the flag is embedded by forcing the least significant bit of the second pixel in the pair. The extraction procedure reads this flag, reconstructs the modified chunk length, and, if $R_i=[L_i,U_i]$7, prepends a 1 to the recovered binary string (Mandal et al., 2012).

This overflow/underflow-aware construction preserves the central APVD principle while adding deterministic side information. Its significance lies in guaranteeing that no stego pixel ever leaves the legal grayscale range, thereby eliminating clipping-induced failure modes that are present in naive PVD-style adjustment.

3. “Unused blocks” and pseudorandom pixel selection

A later development identifies a specific limitation of standard APVD: the “unused blocks” issue. In this account, some pixel pairs cannot safely realize the desired difference adjustment because their original difference lies near a range boundary and the required modification would push one or both pixels outside $R_i=[L_i,U_i]$8. The boundary-checking step then forces a reduction of the payload length for that pair, sometimes to zero. The paper reports that empirically, roughly $R_i=[L_i,U_i]$9–$|R_i|=U_i-L_i+1$0 of pairs fall into such “unused” or “partially-used” cases, which lowers total capacity and introduces a non-uniform embedding pattern exploitable by steganalysis (Hosain et al., 8 Jul 2025).

To mitigate this effect, the paper introduces keyed pseudorandom pixel selection. Let the image contain $|R_i|=U_i-L_i+1$1 non-overlapping pairs. A shared secret seed $|R_i|=U_i-L_i+1$2 initializes a cryptographically-secure PRNG, and the pair index at step $|R_i|=U_i-L_i+1$3 is defined as

$|R_i|=U_i-L_i+1$4

Embedding then proceeds over the pseudorandomly generated pair order rather than strict raster order. For each selected pair, the encoder computes the APVD range, attempts to embed up to $|R_i|=U_i-L_i+1$5 bits, and if a boundary violation occurs, decrements the bit count and retries. Extraction re-generates the same sequence from the same seed and processes pairs in the identical order.

The stated effect is twofold. First, by randomizing the order, previously “unused” pairs are revisited until they can safely embed a smaller number of bits, thereby increasing overall capacity. Second, modifications are spread more uniformly over the image, reducing visually coherent artifacts and weakening fixed spatial correlation patterns. The same paper reports that the pseudorandom revisit of “would-be unused” pairs reduces zero-bit pairs by $|R_i|=U_i-L_i+1$6, raising average bits per pair from $|R_i|=U_i-L_i+1$7 in standard APVD to $|R_i|=U_i-L_i+1$8 (Hosain et al., 8 Jul 2025).

This suggests that the principal novelty is not a change to the APVD embedding law itself, but to the traversal policy and keying model. A plausible implication is that APVD’s security characteristics depend not only on range design and difference remapping but also on the spatial schedule by which pixel pairs are selected.

4. Unified deep-learning steganalysis and reverse steganalysis

APVD has also been studied from the adversarial perspective of steganalysis and reverse steganalysis. A 2025 paper proposes a deep learning-based approach for detecting APVD steganography and reconstructing the hidden payload using a single unified model (Deb et al., 20 Nov 2025). The model is a dual-head Convolutional Neural Network with Squeeze-and-Excitation attention modules.

The reported overall flow is as follows. The input is a single-channel $|R_i|=U_i-L_i+1$9 stego or cover image. A shared feature extractor comprises five convolutional blocks of the form Conv2D $R_i$0 ReLU $R_i$1 BatchNorm $R_i$2 Pooling. Between blocks, SE attention units re-weight channel activations to focus on stego artifacts. The shared representation then branches into two heads. The detection head applies Global Average Pooling, Dense, and Sigmoid to produce $R_i$3. The recovery head applies upsampling blocks and Conv2D with Sigmoid to produce a $R_i$4 probability map of embedded bits.

The hyperparameters reported for this architecture are: input shape $R_i$5; convolutional filters $R_i$6; kernel size $R_i$7; SE ratio $R_i$8; batch size $R_i$9; Adam with learning rate $b = \left\lfloor \log_2 (U_i-L_i+1)\right\rfloor.$0; and $b = \left\lfloor \log_2 (U_i-L_i+1)\right\rfloor.$1 epochs (Deb et al., 20 Nov 2025). The model is trained end-to-end on a combined objective

$b = \left\lfloor \log_2 (U_i-L_i+1)\right\rfloor.$2

with $b = \left\lfloor \log_2 (U_i-L_i+1)\right\rfloor.$3 in the reported experiments. The detection term is binary cross-entropy, and the recovery term is mean squared error on bit-maps. Early stopping is applied on validation loss, and the learning rate decays by $b = \left\lfloor \log_2 (U_i-L_i+1)\right\rfloor.$4 every $b = \left\lfloor \log_2 (U_i-L_i+1)\right\rfloor.$5 epochs without improvement.

The study states that the model was trained and validated on $b = \left\lfloor \log_2 (U_i-L_i+1)\right\rfloor.$6 images from the BOSSbase and UCID datasets. Its purpose is explicitly dual: to identify whether APVD is present and, simultaneously, to perform reverse steganalysis by estimating the embedded payload. In the paper’s framing, this deconstructs an adaptive steganographic scheme that had been valued for high embedding capacity and invisibility.

5. Empirical performance across embedding, image quality, and recovery

The literature reports APVD performance under three different evaluation regimes: fidelity-preserving overflow control, pseudorandomized embedding quality and capacity, and deep-learning-based detection and payload reconstruction.

In the overflow/underflow-aware APVD of Mandal and Das, experiments used the standard PVD range-width set $b = \left\lfloor \log_2 (U_i-L_i+1)\right\rfloor.$7 and evaluated ten standard $b = \left\lfloor \log_2 (U_i-L_i+1)\right\rfloor.$8 test images. The reported result is that APVD achieves exactly the same payload as standard PVD, with PSNR varying by at most $b = \left\lfloor \log_2 (U_i-L_i+1)\right\rfloor.$9 dB relative to PVD. The paper characterizes the visual quality as identical for practical purposes while guaranteeing $d$0 (Mandal et al., 2012).

In the pseudorandom-selection extension, evaluation on eight standard $d$1 images—four grayscale and four color—yielded grayscale PSNRs of approximately $d$2–$d$3 dB, $d$4, and $d$5; for color images, the reported values were approximately $d$6–$d$7 dB, $d$8, and $d$9. Relative to the quantum-S-box method of Abd et al., the paper states that the new method typically improves PSNR by $d$0–$d$1 dB, SSIM by $d$2–$d$3, and UIQ by similar margins (Hosain et al., 8 Jul 2025).

For unified APVD detection and reverse steganalysis, the deep model reports an overall detection accuracy of $d$4, with precision, recall, and F1 of $d$5, $d$6, and $d$7, respectively. The paper also reports that performance drops slightly, by approximately $d$8, at the lowest embedding density of $d$9 bpp. Against an SVM+SPAM baseline, a paired t-test yielded $(P_1,P_2)$00, with clear improvement in both detection and recovery (Deb et al., 20 Nov 2025).

Payload size (bpp)	Bit Error Rate	Recovery Rate
0.2	6.4%	93.6%
0.5	11.7%	88.3%
0.8	17.3%	82.7%

The same reverse-steganalysis study reports a nearly linear increase in BER with payload size, with Pearson $(P_1,P_2)$01 (Deb et al., 20 Nov 2025). This supports the paper’s statement that there is a strong inverse relationship between payload size and recovery accuracy. A common misconception is that adaptive embedding alone is sufficient to prevent systematic detection; the reported CNN results indicate instead that even highly adaptive schemes like APVD leave exploitable fingerprints.

6. Forensic significance, security implications, and prospective directions

The APVD literature presents the method as a high-capacity spatial-domain steganographic framework whose principal design objective is to adapt embedding strength to local image characteristics. The more recent literature also emphasizes that this objective does not eliminate attack surface. The reverse-steganalysis study states that even highly adaptive schemes like APVD leave exploitable fingerprints that a CNN can learn, and that joint detection–recovery models can reconstruct covert messages at useful fidelity, especially at low bpp (Deb et al., 20 Nov 2025).

The forensic implications are described directly. The deep-learning study states that such a system enables law-enforcement to not only flag stego-media but also retrieve hidden evidence, and represents a step towards fully automated reverse steganalysis pipelines. At the same time, the paper notes ethical and legal considerations: while powerful in forensics, reverse-engineering tools also raise privacy concerns, and responsible disclosure with regulated access is described as essential (Deb et al., 20 Nov 2025).

From the defensive side, the papers outline several design responses. One is to pre-encrypt payloads so that, even if decoded by an adversarial network, recovered bits are unintelligible. Another is to consider generative approaches, including GAN-based embedding, to produce images statistically closer to natural priors. A third is to introduce content-aware randomized embedding keys to foil fixed-pattern detectors (Deb et al., 20 Nov 2025). The pseudorandom-selection paper also identifies the quality of the PRNG as a security hinge and proposes future work based on block-cipher-based or hash-based PRNGs with provable unpredictability. Additional directions named there include extension to transform domains such as DCT and wavelets, joint optimization of APVD range tables and PRNG ordering via machine learning, and deeper steganalytic benchmarking against modern deep-learning detectors (Hosain et al., 8 Jul 2025).

Taken together, these developments place APVD at the intersection of three research trajectories: adaptive spatial-domain embedding, keyed randomization for capacity and security management, and AI-driven deconstruction of stego content. The resulting picture is technically specific rather than categorical. APVD remains a method in which capacity is tied to local difference structure, but the contemporary literature shows that its practical security depends on overflow management, pair-selection policy, and the strength of the adversary’s learned detector no less than on the original adaptive embedding rule itself.