Exploiting Vulnerabilities in Speech Translation Systems through Targeted Adversarial Attacks
Anonymous
Here, we provide some related adversarial music/speech samples associated with the paper "Exploiting Vulnerabilities in Speech Translation Systems through Targeted Adversarial Attacks", along with their corresponding translation results. Additionally, we provide a demonstration video showcasing several cases of physical-world attacks.
Adversarial Music Samples
We provide some adversarial music samples here, along with their corresponding translation results.
Attack Seamless Large
| Adversarial Music | Translate to | Translated Text | Corresponding English | Translated Speech |
|---|---|---|---|---|
|
(Target: Are you insane?) |
English(Seen) | "Are you crazy?" | "Are you crazy?" | |
| Mandarin(Seen) | "你是疯了吗?" | "Are you insane?" | ||
| German(Seen) | "Bist du verrückt?" | "Are you insane?" | ||
| French(Unseen) | "Tu es folle ?" | "Are you insane?" | ||
| Italian(Unseen) | "Sei impazzito?" | "Are you insane?" | ||
| Spanish(Unseen) | "¿Estás loco? ¿Qué estás haciendo?" | "Are you insane? What are you doing?" |
| Adversarial Music | Translate to | Translated Text | Corresponding English | Translated Speech |
|---|---|---|---|---|
|
(Target: Who do you think you're talking to?) |
English(Seen) | "Who do you think you're talking to?" | "Who do you think you're talking to?" | |
| Mandarin(Seen) | "你认为你在跟谁说话?" | "Who do you think you're talking to?" | ||
| German(Seen) | "Mit wem meinst du, sprichst du?" | "Who do you think you're talking to?" | ||
| French(Unseen) | "À qui pensez-vous que vous parlez ?" | "Who do you think you're talking to?" | ||
| Italian(Unseen) | "Con chi pensi che stai parlando?" | "Who do you think you're talking to?" | ||
| Spanish(Unseen) | "¿Con quién crees que estás hablando?" | "Who do you think you're talking to?" |
Attack Seamless Expressive
| Adversarial Music | Translate to | Translated Text | Corresponding English | Translated Speech |
|---|---|---|---|---|
|
(Target: Are you insane?) |
English(Seen) | "Are you super dumb?" | "Are you super dumb?" | |
| Mandarin(Seen) | "你疯了吗?" | "Are you insane?" | ||
| German(Seen) | "Bist du verrückt?" | "Are you insane?" | ||
| French(Unseen) | "Êtes-vous en train de vous battre ?" | "Are you fighting?" | ||
| Italian(Unseen) | "Sei impazzito?" | "Are you insane?" | ||
| Spanish(Unseen) | "¿Estás loco, Brendan? " | "Are you insane, Brendan? " |
| Adversarial Music | Translate to | Translated Text | Corresponding English | Translated Speech |
|---|---|---|---|---|
|
(Target: Who do you think you're talking to?) |
English(Seen) | "Who do you think you're talking to?" | "Who do you think you're talking to?" | |
| Mandarin(Seen) | "你认为你在和谁说话?" | "Who do you think you're talking to?" | ||
| German(Seen) | "Mit wem glaubst du das zu sagen?" | "Who do you think you're talking to?" | ||
| French(Unseen) | "Avec qui tu crois que tu es en train de parler ?" | "Who do you think you're talking to?" | ||
| Italian(Unseen) | "Con chi credi di stare parlando?" | "Who do you think you're talking to?" | ||
| Spanish(Unseen) | "¿Con quién crees que estás hablando?" | "Who do you think you're talking to?" |
Attack Seamless M4tv2
| Adversarial Music | Translate to | Translated Text | Corresponding English | Translated Speech |
|---|---|---|---|---|
|
(Target: Are you insane?) |
English(Seen) | "Are you out of your mind?" | "Are you out of your mind?" | |
| Mandarin(Seen) | "你疯了吗?" | "Are you insane?" | ||
| German(Seen) | "Bist du verrückt?" | "Are you insane?" | ||
| French(Unseen) | "Tu es folle ! Tu es folle !" | "You're crazy! You're crazy!" | ||
| Italian(Unseen) | "Puzza di lanto!" | "It stinks of lanto!" | ||
| Spanish(Unseen) | "¿Estás loco? ¿Estás loco? " | "Are you insane? Are you insane?" |
| Adversarial Music | Translate to | Translated Text | Corresponding English | Translated Speech |
|---|---|---|---|---|
|
(Target: Who do you think you're talking to?) |
English(Seen) | "Who do you think you're talking to?" | "Who do you think you're talking to?" | |
| Mandarin(Seen) | "你认为你在跟谁说话?" | "Who do you think you're talking to?" | ||
| German(Seen) | "Mit wem denkst du, sprichst du?" | "Who do you think you're talking to?" | ||
| French(Unseen) | "Tu penses avec qui tu parles?" | "Do you think about who you're talking to?" | ||
| Italian(Unseen) | "Con chi pensi che stai parlando?" | "Who do you think you're talking to?" | ||
| Spanish(Unseen) | "¿Con quién crees que hablas?" | "Who do you think you're talking to?" |
Attack Seamless Medium
| Adversarial Music | Translate to | Translated Text | Corresponding English | Translated Speech |
|---|---|---|---|---|
|
(Target: Are you insane?) |
English(Seen) | "Are you a maniac?" | "Are you a maniac?" | |
| Mandarin(Seen) | "你是疯了吗?" | "Are you insane?" | ||
| German(Seen) | "Sind Sie verrückt?" | "Are you insane?" | ||
| French(Unseen) | "Vous êtes en panne ?" | "Are you out of order?" | ||
| Italian(Unseen) | "Sei un manicista?" | "Are you a manicist?" | ||
| Spanish(Unseen) | "¿Te has vuelto loco?" | "Have you lost your mind?" |
| Adversarial Music | Translate to | Translated Text | Corresponding English | Translated Speech |
|---|---|---|---|---|
|
(Target: Who do you think you're talking to?) |
English(Seen) | "Who do you think you're talking to?" | "Who do you think you're talking to?" | |
| Mandarin(Seen) | "你认为你和谁在说话?" | "Who do you think you're talking to?" | ||
| German(Seen) | "Mit wem glaubst du, du sprichst?" | "Who do you think you're talking to?" | ||
| French(Unseen) | "À qui pensez-vous que vous parlez ?" | "Who do you think you're talking to?" | ||
| Italian(Unseen) | "Con chi credi di parlare?" | "Who do you think you're talking to?" | ||
| Spanish(Unseen) | "¿Con quién crees que estás hablando?" | "Who do you think you're talking to?" |
Attack Canary
| Adversarial Music | Translate to | Translated Text | Corresponding English |
|---|---|---|---|
(Target: Are you insane?) |
English(Seen) | "Are you crazy?" | "Are you crazy?" |
| French(Seen) | "Êtes-vous fou ?" | "Are you insane?" | |
| German(Seen) | "Bist du verrückt?" | "Are you insane?" | |
| Spanish(Unseen) | "¿Estás bien loco?" | "Are you insane?" |
| Adversarial Music | Translate to | Translated Text | Corresponding English |
|---|---|---|---|
(Target: Who do you think you're talking to?) |
English(Seen) | "Who do you think you are talking to?" | "Who do you think you are talking to?" |
| French(Seen) | "A qui pensez-vous que vous parlez ?" | "Who do you think you're talking to?" | |
| German(Seen) | "Mit wem denkst du, dass du redest?" | "Who do you think you're talking to?" | |
| Spanish(Unseen) | "¿A quién crees que estás hablando?" | "Who do you think you're talking to?" |
Adversarial Perturbation Samples
We provide some adversarial speech samples here, along with their corresponding translation results.
Attack Seamless Large
| Adversarial Speech | Translate to | Translated Text | Corresponding English | Translated Speech |
|---|---|---|---|---|
|
ε = 0.01 Original: "He's delighted too with the new premises." Target: "Shame on you." |
English(Seen) | "You should be ashamed of yourself." | "You should be ashamed of yourself." | |
| Mandarin(Seen) | "你应该感到羞耻." | "You should be ashamed." | ||
| German(Seen) | "Schämen Sie sich." | "Shame on you." | ||
| French(Seen) | "Vous devriez avoir honte." | "Shame on you." | ||
| Italian(Unseen) | "Dovresti vergognarti." | "Shame on you." | ||
| Spanish(Unseen) | "¿Te avergonzarías?" | "Would you be ashamed?" |
| Adversarial Speech | Translate to | Translated Text | Corresponding English | Translated Speech |
|---|---|---|---|---|
|
ε = 0.1 Original: "He's delighted too with the new premises." Target: "Don't waste my time anymore." |
English(Seen) | "Don't waste my time anymore." | "Don't waste my time anymore." | |
| Mandarin(Seen) | "不要再浪费我的时间了." | "Don't waste my time anymore." | ||
| German(Seen) | "Verschwenden Sie meine Zeit nicht mehr." | "Don't waste my time anymore." | ||
| French(Seen) | "Ne me fais plus perdre mon temps." | "Don't waste my time anymore." | ||
| Italian(Unseen) | "Non sprecare più il mio tempo." | "Don't waste my time anymore." | ||
| Spanish(Unseen) | "No vuelvas a desperdiciar mi tiempo." | "Don't waste my time again." |
| Adversarial Speech | Translate to | Translated Text | Corresponding English | Translated Speech |
|---|---|---|---|---|
|
ε = 0.5 Original: "He's delighted too with the new premises." Target: "Who do you think you're talking to?" |
English(Seen) | "Who do you think you're talking to?" | "Who do you think you're talking to?" | |
| Mandarin(Seen) | "你认为你在跟谁说话?" | "Who do you think you're talking to?" | ||
| German(Seen) | "Mit wem glaubst du, sprichst du?" | "Who do you think you're talking to?" | ||
| French(Seen) | "À qui pensez-vous que vous parlez ?" | "Who do you think you're talking to?" | ||
| Italian(Unseen) | "Con chi pensi che stai parlando?" | "Who do you think you're talking to?" | ||
| Spanish(Unseen) | "¿Con quién crees que estás hablando?" | "Who do you think you're talking to?" |
Attack Seamless Expressive
| Adversarial Speech | Translate to | Translated Text | Corresponding English | Translated Speech |
|---|---|---|---|---|
|
ε = 0.01 Original: "He's delighted too with the new premises." Target: "Shame on you." |
English(Seen) | "You should be ashamed of yourself." | "You should be ashamed of yourself." | |
| Mandarin(Seen) | "你们这两个人可可耻了." | "Shame on you two." | ||
| German(Seen) | "Du solltest dich schämen, dass du dich so verärgert hast." | "You should be ashamed of yourself for being so upset." | ||
| French(Seen) | "Vous devriez être honteux de votre méchanceté." | "You should be ashamed of your wickedness." | ||
| Italian(Unseen) | "Ti vergogneresti di te stesso." | "You'd be ashamed of yourself." | ||
| Spanish(Unseen) | "Deberías estar avergonzado de tus malas premisas." | "You should be ashamed of your bad premises." |
| Adversarial Speech | Translate to | Translated Text | Corresponding English | Translated Speech |
|---|---|---|---|---|
|
ε = 0.1 Original: "He's delighted too with the new premises." Target: "Who do you think you're talking to?" |
English(Seen) | "Who do you think you're talking to?" | "Who do you think you're talking to?" | |
| Mandarin(Seen) | "你认为你在和谁说话? " | "Who do you think you're talking to?" | ||
| German(Seen) | "Mit wem denkst du, sprichst du mit?" | "Who do you think you're talking to?" | ||
| French(Seen) | "Tu crois que tu parles à qui ?" | "Who do you think you're talking to?" | ||
| Italian(Unseen) | "Con chi pensi che stia parlando?" | "Who do you think he's talking to?" | ||
| Spanish(Unseen) | "¿Con quién crees que estás hablando?" | "Who do you think you're talking to?" |
| Adversarial Speech | Translate to | Translated Text | Corresponding English | Translated Speech |
|---|---|---|---|---|
|
ε = 0.5 Original: "He's delighted too with the new premises." Target: "Are you insane?" |
English(Seen) | "Are you insane, please?" | "Are you insane, please?" | |
| Mandarin(Seen) | "你疯了吗? " | "Are you crazy?" | ||
| German(Seen) | "Bist du verrückt, oder bist du verrückt?" | "Are you crazy, or are you crazy?" | ||
| French(Seen) | "Vous êtes fou, s'il vous plaît ?" | "Are you crazy, please?" | ||
| Italian(Unseen) | "Sei impazzito, ti prego?" | "Are you crazy, please?" | ||
| Spanish(Unseen) | "¿Estás loco? ¿Estás loco por favor?" | "Are you crazy? Are you crazy please?" |
Attack Seamless M4tv2
| Adversarial Speech | Translate to | Translated Text | Corresponding English | Translated Speech |
|---|---|---|---|---|
|
ε = 0.01 Original: "He's delighted too with the new premises." Target: "Who do you think you're talking to?" |
English(Seen) | "Who do you think you're talking to?" | "Who do you think you're talking to?" | |
| Mandarin(Seen) | "他选择了你和谁说话?" | "Who did he choose you to talk to?" | ||
| German(Seen) | "Mit wem soll er jetzt sprechen?" | "Who is he supposed to talk to now?" | ||
| French(Seen) | "Vous êtes sûre de qui vous parlez ?" | "Are you sure who you are talking about?" | ||
| Italian(Unseen) | "Con chi sta parlando, secondo lei?" | "Who do you think he's talking to?" | ||
| Spanish(Unseen) | "¿A quién crees que estás hablando?" | "Who do you think you're talking to? " |
| Adversarial Speech | Translate to | Translated Text | Corresponding English | Translated Speech |
|---|---|---|---|---|
|
ε = 0.1 Original: "He's delighted too with the new premises." Target: "Shame on you." |
English(Seen) | "You should be ashamed." | "You should be ashamed." | |
| Mandarin(Seen) | "你应该感到羞愧." | "You should be ashamed." | ||
| German(Seen) | "Scham dich um dich." | "Shame on you." | ||
| French(Seen) | "Vous devriez avoir honte de vous." | "Shame on you." | ||
| Italian(Unseen) | "Si dovrebbe vergognarvi." | "Shame on you." | ||
| Spanish(Unseen) | "Usted debe sentirse avergonzado." | "You must feel ashamed." |
| Adversarial Speech | Translate to | Translated Text | Corresponding English | Translated Speech |
|---|---|---|---|---|
|
ε = 0.5 Original: "He's delighted too with the new premises." Target: "Are you insane?" |
English(Seen) | "Are you insane?" | "Are you insane?" | |
| Mandarin(Seen) | "你疯了吗?" | "Are you insane?" | ||
| German(Seen) | "Bist du verrückt?" | "Are you insane?" | ||
| French(Seen) | "Vous êtes dingue ?" | "Are you insane?" | ||
| Italian(Unseen) | "Sei un pazzo?" | "Are you insane?" | ||
| Spanish(Unseen) | "He reaccionado también, ¿estás loco?" | "I reacted too, are you crazy?" |
Attack Seamless Medium
| Adversarial Speech | Translate to | Translated Text | Corresponding English | Translated Speech |
|---|---|---|---|---|
|
ε = 0.01 Original: "He's delighted too with the new premises." Target: "Who do you think you're talking to?" |
English(Seen) | "Who do you think you're talking to?" | "Who do you think you're talking to?" | |
| Mandarin(Seen) | "你认为你和谁说话?" | "Who do you think you're talking to?" | ||
| German(Seen) | "Mit wem glaubst du, dass du sprichst?" | "Who do you think you're talking to?" | ||
| French(Seen) | "Avec qui pensez-vous que vous parlez ?" | "Who do you think you're talking to?" | ||
| Italian(Unseen) | "Con chi credi di parlare?" | "Who do you think you're talking to?" | ||
| Spanish(Unseen) | "¿Con quién crees que hablas?" | "Who do you think you're talking to? " |
| Adversarial Speech | Translate to | Translated Text | Corresponding English | Translated Speech |
|---|---|---|---|---|
|
ε = 0.1 Original: "He's delighted too with the new premises." Target: "Are you insane?" |
English(Seen) | "Are you a lunatic?" | "Are you a lunatic?" | |
| Mandarin(Seen) | "你是疯了吗?" | "Are you crazy?" | ||
| German(Seen) | "Sind Sie verrückt?" | "Are you out of your mind?" | ||
| French(Seen) | "Est-ce que vous êtes fou ?" | "Are you crazy?" | ||
| Italian(Unseen) | "Sei pazzo?" | "Are you out of your mind?" | ||
| Spanish(Unseen) | "¿Estás loco? ¿Te ves?" | "Are you crazy? Do you see?" |
| Adversarial Speech | Translate to | Translated Text | Corresponding English | Translated Speech |
|---|---|---|---|---|
|
ε = 0.5 Original: "He's delighted too with the new premises." Target: "You make me sick." |
English(Seen) | "You're making me sick." | "You're making me sick." | |
| Mandarin(Seen) | "你让我生病了." | "You made me sick." | ||
| German(Seen) | "Du machst mich krank." | "You are making me sick." | ||
| French(Seen) | "Vous me mettez malade." | "You are making me sick." | ||
| Italian(Unseen) | "Tu mi metti in malattia." | "You make me sick." | ||
| Spanish(Unseen) | "Tú me haces enfermo." | "You make me sick." |
Attack Canary
| Adversarial Speech | Translate to | Translated Text | Corresponding English |
|---|---|---|---|
|
ε = 0.01 Original: "He's delighted too with the new premises." Target: "You make me sick." |
English(Seen) | "You make me ill." | "You make me ill." |
| French(Seen) | "Tu me rends malade." | "You are making me sick." | |
| German(Seen) | "Du machst mich krank." | "You are making me sick." | |
| Spanish(Unseen) | "Me haces enfermar." | "You make me sick." |
| Adversarial Speech | Translate to | Translated Text | Corresponding English |
|---|---|---|---|
|
ε = 0.1 Original: "He's delighted too with the new premises." Target: "Who do you think you're talking to?" |
English(Seen) | "Who do you think you're talking to?" | "Who do you think you're talking to?" |
| French(Seen) | "À qui pensez-vous que vous parlez ?" | "Who do you think you're talking to?" | |
| German(Seen) | "Mit wem denkst du, dass du redest?" | "Who do you think you're talking to?" | |
| Spanish(Unseen) | "¿Con quién crees que estás hablando?" | "Who do you think you're talking to?" |
| Adversarial Speech | Translate to | Translated Text | Corresponding English |
|---|---|---|---|
|
ε = 0.5 Original: "He's delighted too with the new premises." Target: "Don't waste my time anymore." |
English(Seen) | "Don't waste my time" | "Don't waste my time." |
| French(Seen) | "Ne perdez plus mon temps." | "Don't waste my time anymore." | |
| German(Seen) | "Verschwende meine Zeit nicht mehr." | "Don't waste my time anymore." | |
| Spanish(Unseen) | "No pierdas mi tiempo." | "Don't waste my time." |
Physical Attack Demo
In the over-the-air physical test experiments presented in the paper, we generate six adversarial music samples for each attack and conduct repeated testing to ensure stability. This results in the testing of 60 adversarial music samples for each target language. Additionally, we further present a video here demonstrating a case of physical-world attacks. In a meeting room with dimensions of 5.2m × 7.5m × 3.5m, we accessed the speech translation service via a laptop and used its built-in microphone for audio recording. For the adversarial music, we employed a smartphone to play the audio, thereby injecting malicious content. For Canary, we conducted tests using the official translation service. For Seamless Large, we followed the official instructions from the Seamless developers and deployed it locally using the method described in this link to build a translation server for testing.
Other comments
The proposed adversarial examples are intended for research purposes only and will not be used for any malicious activities.