John X. Morris

John X. Morris, Volodymyr Kuleshov, Vitaly Shmatikov et al.

How much private information do text embeddings reveal about the original text? We investigate the problem of embedding \textit{inversion}, reconstructing the full text represented in dense text embeddings. We frame the problem as controlled generation: generating text that, when reembedded, is close to a fixed point in latent space. We find that although a naïve model conditioned on the embedding performs poorly, a multi-step method that iteratively corrects and re-embeds text is able to recover $92\%$ of $32\text{-token}$ text inputs exactly. We train our model to decode text embeddings from two state-of-the-art embedding models, and also show that our model can recover important personal information (full names) from a dataset of clinical notes. Our code is available on Github: \href{https://github.com/jxmorris12/vec2text}{github.com/jxmorris12/vec2text}.

John X. Morris, Wenting Zhao, Justin T. Chiu et al.

Language models produce a distribution over the next token; can we use this information to recover the prompt tokens? We consider the problem of language model inversion and show that next-token probabilities contain a surprising amount of information about the preceding text. Often we can recover the text in cases where it is hidden from the user, motivating a method for recovering unknown prompts given only the model's current distribution output. We consider a variety of model access scenarios, and show how even without predictions for every token in the vocabulary we can recover the probability vector through search. On Llama-2 7b, our inversion method reconstructs prompts with a BLEU of $59$ and token-level F1 of $78$ and recovers $27\%$ of prompts exactly. Code for reproducing all experiments is available at http://github.com/jxmorris12/vec2text.

Chandan Singh, John X. Morris, Jyoti Aneja et al. · berkeley

Large language models (LLMs) have displayed an impressive ability to harness natural language to perform complex tasks. In this work, we explore whether we can leverage this learned ability to find and explain patterns in data. Specifically, given a pre-trained LLM and data examples, we introduce interpretable autoprompting (iPrompt), an algorithm that generates a natural-language string explaining the data. iPrompt iteratively alternates between generating explanations with an LLM and reranking them based on their performance when used as a prompt. Experiments on a wide range of datasets, from synthetic mathematics to natural-language understanding, show that iPrompt can yield meaningful insights by accurately finding groundtruth dataset descriptions. Moreover, the prompts produced by iPrompt are simultaneously human-interpretable and highly effective for generalization: on real-world sentiment classification datasets, iPrompt produces prompts that match or even improve upon human-written prompts for GPT-3. Finally, experiments with an fMRI dataset show the potential for iPrompt to aid in scientific discovery. All code for using the methods and data here is made available on Github.

John X. Morris, Chandan Singh, Alexander M. Rush et al. · allen-ai

Prompting language models (LMs) is the main interface for applying them to new tasks. However, for smaller LMs, prompting provides low accuracy compared to gradient-based finetuning. Tree Prompting is an approach to prompting which builds a decision tree of prompts, linking multiple LM calls together to solve a task. At inference time, each call to the LM is determined by efficiently routing the outcome of the previous call using the tree. Experiments on classification datasets show that Tree Prompting improves accuracy over competing methods and is competitive with fine-tuning. We also show that variants of Tree Prompting allow inspection of a model's decision-making process.

John X. Morris, Justin T. Chiu, Ramin Zabih et al.

Deidentification seeks to anonymize textual data prior to distribution. Automatic deidentification primarily uses supervised named entity recognition from human-labeled data points. We propose an unsupervised deidentification method that masks words that leak personally-identifying information. The approach utilizes a specially trained reidentification model to identify individuals from redacted personal documents. Motivated by K-anonymity based privacy, we generate redactions that ensure a minimum reidentification rank for the correct profile of the document. To evaluate this approach, we consider the task of deidentifying Wikipedia Biographies, and evaluate using an adversarial reidentification metric. Compared to a set of unsupervised baselines, our approach deidentifies documents more completely while removing fewer words. Qualitatively, we see that the approach eliminates many identifying aspects that would fall outside of the common named entity based approach.

Zach Nussbaum, John X. Morris, Brandon Duderstadt et al.

This technical report describes the training of nomic-embed-text-v1, the first fully reproducible, open-source, open-weights, open-data, 8192 context length English text embedding model that outperforms both OpenAI Ada-002 and OpenAI text-embedding-3-small on the short-context MTEB benchmark and the long context LoCo benchmark. We release the training code and model weights under an Apache 2.0 license. In contrast with other open-source models, we release the full curated training data and code that allows for full replication of nomic-embed-text-v1. You can find code and data to replicate the model at https://github.com/nomic-ai/contrastors.

Tingwei Zhang, Collin Zhang, John X. Morris et al.

We introduce a new type of indirect, cross-modal injection attacks against visual language models that enable creation of self-interpreting images. These images contain hidden "meta-instructions" that control how models answer users' questions about the image and steer models' outputs to express an adversary-chosen style, sentiment, or point of view. Self-interpreting images act as soft prompts, conditioning the model to satisfy the adversary's (meta-)objective while still producing answers based on the image's visual content. Meta-instructions are thus a stronger form of prompt injection. Adversarial images look natural and the model's answers are coherent and plausible, yet they also follow the adversary-chosen interpretation, e.g., political spin, or even objectives that are not achievable with explicit text instructions. We evaluate the efficacy of self-interpreting images for a variety of models, interpretations, and user prompts. We describe how these attacks could cause harm by enabling creation of self-interpreting content that carries spam, misinformation, or spin. Finally, we discuss defenses.

John X. Morris

Adversarial example generation methods in NLP rely on models like language models or sentence encoders to determine if potential adversarial examples are valid. In these methods, a valid adversarial example fools the model being attacked, and is determined to be semantically or syntactically valid by a second model. Research to date has counted all such examples as errors by the attacked model. We contend that these adversarial examples may not be flaws in the attacked model, but flaws in the model that determines validity. We term such invalid inputs second-order adversarial examples. We propose the constraint robustness curve and associated metric ACCS as tools for evaluating the robustness of a constraint to second-order adversarial examples. To generate this curve, we design an adversarial attack to run directly on the semantic similarity models. We test on two constraints, the Universal Sentence Encoder (USE) and BERTScore. Our findings indicate that such second-order examples exist, but are typically less common than first-order adversarial examples in state-of-the-art models. They also indicate that USE is effective as constraint on NLP adversarial examples, while BERTScore is nearly ineffectual. Code for running the experiments in this paper is available at https://github.com/jxmorris12/second-order-adversarial-examples.

John X. Morris, Jin Yong Yoo, Yanjun Qi

TextAttack is an open-source Python toolkit for adversarial attacks, adversarial training, and data augmentation in NLP. TextAttack unites 15+ papers from the NLP adversarial attack literature into a single framework, with many components reused across attacks. This framework allows both researchers and developers to test and study the weaknesses of their NLP models. To build such an open-source NLP toolkit requires solving some common problems: How do we enable users to supply models from different deep learning frameworks? How can we build tools to support as many different datasets as possible? We share our insights into developing a well-written, well-documented NLP Python framework in hope that they can aid future development of similar packages.

Jin Yong Yoo, John X. Morris, Eli Lifland et al.

We study the behavior of several black-box search algorithms used for generating adversarial examples for natural language processing (NLP) tasks. We perform a fine-grained analysis of three elements relevant to search: search algorithm, search space, and search budget. When new search algorithms are proposed in past work, the attack search space is often modified alongside the search algorithm. Without ablation studies benchmarking the search algorithm change with the search space held constant, one cannot tell if an increase in attack success rate is a result of an improved search algorithm or a less restrictive search space. Additionally, many previous studies fail to properly consider the search algorithms' run-time cost, which is essential for downstream tasks like adversarial training. Our experiments provide a reproducible benchmark of search algorithms across a variety of search spaces and query budgets to guide future research in adversarial NLP. Based on our experiments, we recommend greedy attacks with word importance ranking when under a time constraint or attacking long inputs, and either beam search or particle swarm optimization otherwise. Code implementation shared via https://github.com/QData/TextAttack-Search-Benchmark

John X. Morris, Eli Lifland, Jin Yong Yoo et al.

While there has been substantial research using adversarial attacks to analyze NLP models, each attack is implemented in its own code repository. It remains challenging to develop NLP attacks and utilize them to improve model performance. This paper introduces TextAttack, a Python framework for adversarial attacks, data augmentation, and adversarial training in NLP. TextAttack builds attacks from four components: a goal function, a set of constraints, a transformation, and a search method. TextAttack's modular design enables researchers to easily construct attacks from combinations of novel and existing components. TextAttack provides implementations of 16 adversarial attacks from the literature and supports a variety of models and datasets, including BERT and other transformers, and all GLUE tasks. TextAttack also includes data augmentation and adversarial training modules for using components of adversarial attacks to improve model accuracy and robustness. TextAttack is democratizing NLP: anyone can try data augmentation and adversarial training on any model or dataset, with just a few lines of code. Code and tutorials are available at https://github.com/QData/TextAttack.

Collin Zhang, John X. Morris, Vitaly Shmatikov

We consider the problem of language model inversion: given outputs of a language model, we seek to extract the prompt that generated these outputs. We develop a new black-box method, output2prompt, that learns to extract prompts without access to the model's logits and without adversarial or jailbreaking queries. In contrast to previous work, output2prompt only needs outputs of normal user queries. To improve memory efficiency, output2prompt employs a new sparse encoding techique. We measure the efficacy of output2prompt on a variety of user and system prompts and demonstrate zero-shot transferability across different LLMs.

Wilson Wu, John X. Morris, Lionel Levine

Do transformers "think ahead" during inference at a given position? It is known transformers prepare information in the hidden states of the forward pass at time step $t$ that is then used in future forward passes $t+τ$. We posit two explanations for this phenomenon: pre-caching, in which off-diagonal gradient terms present during training result in the model computing features at $t$ irrelevant to the present inference task but useful for the future, and breadcrumbs, in which features most relevant to time step $t$ are already the same as those that would most benefit inference at time $t+τ$. We test these hypotheses by training language models without propagating gradients to past timesteps, a scheme we formalize as myopic training. In a constructed synthetic data setting, we find clear evidence for pre-caching. In the autoregressive language modeling setting, our experiments are more suggestive of the breadcrumbs hypothesis, though pre-caching increases with model scale.

Rishi Jha, Collin Zhang, Vitaly Shmatikov et al.

We introduce the first method for translating text embeddings from one vector space to another without any paired data, encoders, or predefined sets of matches. Our unsupervised approach translates any embedding to and from a universal latent representation (i.e., a universal semantic structure conjectured by the Platonic Representation Hypothesis). Our translations achieve high cosine similarity across model pairs with different architectures, parameter counts, and training datasets. The ability to translate unknown embeddings into a different space while preserving their geometry has serious implications for the security of vector databases. An adversary with access only to embedding vectors can extract sensitive information about the underlying documents, sufficient for classification and attribute inference.

John X. Morris, Chawin Sitawarin, Chuan Guo et al. · deepmind, meta-ai

We propose a new method for estimating how much a model knows about a datapoint and use it to measure the capacity of modern language models. Prior studies of language model memorization have struggled to disentangle memorization from generalization. We formally separate memorization into two components: unintended memorization, the information a model contains about a specific dataset, and generalization, the information a model contains about the true data-generation process. When we completely eliminate generalization, we can compute the total memorization, which provides an estimate of model capacity: our measurements estimate that GPT-style models have a capacity of approximately 3.6 bits per parameter. We train language models on datasets of increasing size and observe that models memorize until their capacity fills, at which point "grokking" begins, and unintended memorization decreases as models begin to generalize. We train hundreds of transformer language models ranging from $500K$ to $1.5B$ parameters and produce a series of scaling laws relating model capacity and data size to membership inference.

Murtaza Nazir, Matthew Finlayson, John X. Morris et al.

Language model inversion seeks to recover hidden prompts using only language model outputs. This capability has implications for security and accountability in language model deployments, such as leaking private information from an API-protected language model's system message. We propose a new method -- prompt inversion from logprob sequences (PILS) -- that recovers hidden prompts by gleaning clues from the model's next-token probabilities over the course of multiple generation steps. Our method is enabled by a key insight: The vector-valued outputs of a language model occupy a low-dimensional subspace. This enables us to losslessly compress the full next-token probability distribution over multiple generation steps using a linear map, allowing more output information to be used for inversion. Our approach yields massive gains over previous state-of-the-art methods for recovering hidden prompts, achieving 2--3.5 times higher exact recovery rates across test sets, in one case increasing the recovery rate from 17% to 60%. Our method also exhibits surprisingly good generalization behavior; for instance, an inverter trained on 16 generations steps gets 5--27 points higher prompt recovery when we increase the number of steps to 32 at test time. Furthermore, we demonstrate strong performance of our method on the more challenging task of recovering hidden system messages. We also analyze the role of verbatim repetition in prompt recovery and propose a new method for cross-family model transfer for logit-based inverters. Our findings show that next-token probabilities are a considerably more vulnerable attack surface for inversion attacks than previously known.

Collin Zhang, John X. Morris, Vitaly Shmatikov

Embedding inversion, i.e., reconstructing text given its embedding and black-box access to the embedding encoder, is a fundamental problem in both NLP and security. From the NLP perspective, it helps determine how much semantic information about the input is retained in the embedding. From the security perspective, it measures how much information is leaked by vector databases and embedding-based retrieval systems. State-of-the-art methods for embedding inversion, such as vec2text, have high accuracy but require (a) training a separate model for each embedding, and (b) a large number of queries to the corresponding encoder. We design, implement, and evaluate ZSInvert, a zero-shot inversion method based on the recently proposed adversarial decoding technique. ZSInvert is fast, query-efficient, and can be used for any text embedding without training an embedding-specific inversion model. We measure the effectiveness of ZSInvert on several embeddings and demonstrate that it recovers key semantic information about the corresponding texts.

John X. Morris, Junjie Oscar Yin, Woojeong Kim et al.

Modern language models often have open weights but closed training data. We formalize the problem of data approximation from model weights and propose several baselines and metrics. We develop a gradient-based approach that selects the highest-matching data from a large public text corpus and show its effectiveness at recovering useful data given only weights of the original and finetuned models. Even when none of the true training data is known, our method is able to locate a small subset of public Web documents can be used to train a model to close to the original model performance given models trained for both classification and supervised-finetuning. On the AG News classification task, our method improves performance from 65% (using randomly selected data) to 80%, approaching the expert benchmark of 88%. When applied to a model trained with SFT on MSMARCO web documents, our method reduces perplexity from 3.3 to 2.3, compared to an expert LLAMA model's perplexity of 2.0.

John X. Morris, Eli Lifland, Jack Lanchantin et al.

State-of-the-art attacks on NLP models lack a shared definition of a what constitutes a successful attack. We distill ideas from past work into a unified framework: a successful natural language adversarial example is a perturbation that fools the model and follows some linguistic constraints. We then analyze the outputs of two state-of-the-art synonym substitution attacks. We find that their perturbations often do not preserve semantics, and 38% introduce grammatical errors. Human surveys reveal that to successfully preserve semantics, we need to significantly increase the minimum cosine similarities between the embeddings of swapped words and between the sentence encodings of original and perturbed sentences.With constraints adjusted to better preserve semantics and grammaticality, the attack success rate drops by over 70 percentage points.