TL;DR
Multimodal text generation models are often assumed to be robust because one modality can compensate for errors in another. This paper introduces Iterative-attack, the first multimodal adversarial attack framework that breaks this defense by iteratively coordinating perturbations across both text and image inputs, exposing significant vulnerabilities in state-of-the-art models.
What Problem Does This Paper Address?
Deep neural networks are notoriously vulnerable to adversarial attacks—small, imperceptible changes to inputs that cause models to fail catastrophically. While single-modality attacks (altering just text or just images) are well-studied, multimodal text generation tasks like Image-guided Story Ending Generation (IgSEG) present a unique challenge.
In these tasks, if an attacker only perturbs the text, the unperturbed image provides complementary information that "corrects" the model's understanding, causing the attack to fail (and vice versa).
The authors tackle the challenge of finding the most vulnerable cross-modal information patches across discrete text spaces and continuous image spaces simultaneously, preventing the modalities from correcting each other.
Methods
The proposed Iterative-attack method operates in two main steps to bridge the gap between discrete text tokens and continuous image pixels:
- Potential Adversarial Text Generation: The algorithm identifies the most influential words in the story context using an importance score based on a masking strategy. For these critical words, it generates candidate substitutions combining character-level typos ("bugs") and word-level synonyms via BERT.
- Iterative Multimodal Optimization: Instead of attacking modalities independently, the method fuses them. For each text substitution, it applies a continuous image perturbation using Projected Gradient Descent (PGD). The optimization targets an adversarial loss function ($L_{adv}$) designed to minimize the probability of the model generating the correct next token.
This coordinated loop continues until the story ending quality drops below a predefined threshold.
Datasets
The authors evaluated their approach on three benchmark datasets:
- VIST-E: Derived from the Visual Storytelling (VIST) dataset, containing 49,913 training, 4,963 validation, and 5,030 test samples. Each sample includes a 4-sentence story context, an ending-related image, and a ground-truth story ending.
- LSMDC-E: Derived from the Large Scale Movie Description Challenge (LSMDC) 2021 dataset, consisting of 20,151 training, 1,477 validation, and 2,005 test samples.
- Multi30k: A multimodal machine translation dataset used to evaluate the generalizability of the attack beyond story generation.
Key Insights and Ideas
- The Correction Dilemma: Single-modality attacks suffer from a bottleneck where the unperturbed modality acts as a fail-safe. True multimodal vulnerability lies in the interaction between modalities.
- Attention Misdirection: Visualization via Grad-CAM reveals that successful attacks don't just change the text; they completely shift the target model's visual attention map, forcing it to focus on irrelevant image regions while keeping the image perturbations entirely imperceptible to humans.
- The Discrete-Continuous Synergy: Alternating between discrete word swaps and continuous pixel gradients allows the attack to find highly effective adversarial "patch pairs" that standard step-wise or single-modality methods miss.
Key Results
The authors tested their attack against four target architectures: Seq2Seq, Transformer, MGCL, and MMT.
- High Success Rates: On the VIST-E dataset, Iterative-attack achieved an Attack Success Rate (ASR) of 57.09% against Seq2Seq and 39.67% against advanced Multimodal Memory Transformers (MMT). This vastly outperformed text-only baselines like kNN and CharSwap, which hovered around 12–31%.
- Semantic Preservation: Despite drastically degrading generation quality (reducing BLEU scores by over 30–50%), the adversarial text maintained a high semantic similarity (>0.95) to the original text, ensuring the attack remains stealthy to human observers.
- Cross-Task Generalizability: When tested on the Multi30k dataset for machine translation, the method maintained its superiority, achieving an ASR of ~45% compared to ~20% for non-iterative multimodal attacks (Co-attack).
Limitations and Future Work
The attack's efficacy is inherently bounded by the target model's baseline capability. Through error analysis, the authors noted that if a model already ignores or fails to comprehend the input image effectively on clean data, perturbing the image yields little to no additional disruptive power.
Future work will focus on developing robust defense mechanisms (such as multimodal adversarial training) to counter these coordinated cross-modal attacks and establishing a more comprehensive evaluation benchmark for multimodal text generation robustness.
Summary updated